This paper presents and integrated view of personalized information spaces. The topics that are described cover different dimensions of describing, customizing, reusing and presenting multimedia content. In order to have personalized multimedia content the systems should provide the ability to access it when and where it is necessary, in a way that is appropriate for each specific user. The paper describes several multimedia information processing and visualization systems that explore these concepts. More specifically, the applications and systems include annotation tools for describing and enhancing multimedia content, personalization and customization techniques, and tools for anytime, anywhere access to multimedia information. The examples of information access are an augmented reality system and applications for mobile devices, representing two of the dominant and convergent trends in ubiquitous computing.

With the recent advances in computer technology, medical images and multimedia information have a major impact to our modem life. This Keynote presentation will give a brief introduction of the current research on medical images and multimedia data management and processing, conducted at the Center for Multimedia Signal Processing (CMSP) and the Biomedical and Multimedia Information Technology (BMIT) Group, as well as in other major research laboratories.

In non-parametric pattern recognition, the probability density function is approximated by means of many parameters, each one for a density value in a small hyper-rectangular volume of the space. The hyper-rectangles are determined by appropriately quantizing the range of each variable. Optimal quantization determines a compact and efficient representation of the probability density of data by optimizing a global quantizer performance measure. The measure used here is a weighted combination of average log likelihood, entropy and correct classification probability. In multi-dimensions, we study a grid based quantization technique. Smoothing is an important aspect of optimal quantization because it affects the generalization ability of the quantized density estimates. We use a fast generalized k nearest neighbor smoothing algorithm. We illustrate the effectiveness of optimal quantization on a set of not very well separated Gaussian mixture models, as compared to the expectation maximization (EM) algorithm. Optimal quantization produces better results than the EM algorithm. The reason is that the convergence of the EM algorithm to the true parameters for not well separated mixture models can be extremely slow.

Despite research activity during the past decade, the problem of how to carry out 3D scene reconstruction from a video sequence is not fully solved. Many techniques exist, useful in different situations. However, for a general scene, methods can still be quite costly in terms of time. The present paper discusses methods whereby the existence of planar or almost planar sections of the scene may be exploited throughout the reconstruction process. We consider the homography induced by a plane to aid in point tracking, projective reconstruction, self calibration and model building. Two recent algorithms for projective reconstruction in the presence of planes are reexamined and presented in a new form.

An algorithm complexity is a very crucial issue in the algorithm design, especially if large data sets are to be processed. Data search is very often used in many algorithms and hash function use gives us a possibility to speed up the process significantly. Nevertheless, it is very difficult do design a good hash function especially for geometric applications. This paper describes a new hash function, its behaviour and use for non-trivial problems. Some problems can be solved effectively using the principle of duality and the hash data structure. Also some problems that cannot be solved in Euclidean space can be solved if dual representation is used and some examples are presented, too.

We are participating in the international competition to develop robots that can play football (or soccer as it is known in the US and Canada). The competition consists of several leagues each of which examines a different technology but shares the common goal of advancing the skills of autonomous robots, robots that function without a central hierarchical command structure. The Dutch team, Clockwork Orange, involves several universities and the contribution of our group at the TU Deift is in the domain of robot vision and motion. In this paper we will describe the background to the project, the characteristics of the robots in our league, our approach to various vision tasks, their implementation in computer architectures, and the results of our efforts.

Image compression based on regions of interest (ROl) means to compress interesting regions in an image with high quality, and to compress uninteresting regions with relatively low quality. Based on this idea, a multi-threshold fractal image-coding algorithm based on regions of interest is proposed in this paper. It uses different error threshold for different regions, and puts both near-lossless coding in the ROl and lossy coding in the UROI (uninteresting regions) under the same fractal frame. Quadtree partition algorithm is employed to compresses the regions of interest near-losslessly and the other regions roughly. By using this algorithm, good decoding image quality of the regions of interest can be obtained while maintaining high compression ratio. Image coding time is also shortened greatly. Simulation results for some medical images have shown the effectiveness of the proposed method. As compared to other known method, the proposed method is very attractive both in computation and in storage.

Previous error control methods have obvious defects more or less for the videoconference applications. A simple and efficient error control mechanism based on adaptive intra-frame refreshment, in this paper, is proposed to conceal the transmission errors ofreal time video. This scheme maintains a backward channel to feedback the error information found by the decoder so that the coder is able to adjust the interval of intra-frame coding dynamically to conceal all sorts of real time video transmission errors. Simulation results present that this method can efficiently eliminate the temporal and spatial error propagation. Compared to conventional schemes, our algorithm need not modify the coder and decoder (Codec) and no more CPU time is consumed as well. As a result, this algorithm is especially fit for the videoconference or visual telephone applications.

In this paper, a novel robust image coder with scalable resolution is presented, called Robust ZeroBlock Wavelet (RZBW), which is suitable for image transport over a noisy channel. In the coder, the zeroblock-based coding algorithm is used, which proved to be an efficient technique for exploiting the clustering of energy found in image transforms. The coder provides both excellent compression performance and graceful degradation over noisy channel. The coder compresses the wavelet coefficients from low frequency to high frequency, so the resolutions ofthe reconsiructed image are scalable.

MPEG-4 is applied to varieties of current and future multimedia applications. The standard not only supports existed frame-based coding standards such as MPEG-i, MPEG-2 or H.263, but also provides content-based representation and flexible toolbox, which lead to MPEG-4 more complicated. This paper first briefly presents the implementation method of the video decoding of MPEG-4 Core Visual Profile, which is a subset of MPEG-4 standard. The Core Visual Profile is quite suitable to streaming video, which will possibly become the hot spot for the development ofMPEG-4. Then, the paper proposes a design scheme for the basic hardware structure of the decoding system based on TMS32OC6X DSP, and simply analyzes the decoding processing of the system.

A novel stereoscopic coding-decoding framework was proposed in this paper. The reference image stream of a stereoscopic sequence is independently coded by a conventional MPEG-type scheme at higher quality; only a few reference frames in target stream are coded and transmitted. The rest frames are 'skipped' and are reconstructed at the decoder using a Combine Stereoscopic Frame Estimation and Interpolation without search (CSFEI) technique proposed in this paper. In order to estimate the unfilled regions due to occlusion, a global solution based on a probability model is also proposed. Simulation results show that the CSFEI scheme can achieve a PSNR gain (about 2-3 dB) as compared to the block-based matching algorithm.

In this paper, we pay attention to the application of image compression and reconstruction with Newton-Thiele's rational interpolation theorem . We present a new algorithm which is suited for fast computation with less distortion . Many examples are given, and the result shows that this algorithm is more convenient in computation and more efficient in implementing reconstruction.

In this paper, the properties of the gray level mixing function in two-dimensional chaotic map image encryption are analyzed. We also discuss the necessity of introducing parameters and the methods how to present diffusion mechanism and interfuse a pseudo-random number sequence in the gray level mixing function. We proposed a new kind of gray level mixing function which not only have diffusion mechanism but also have random variable . It is shown in experiment that this method has good performance.

A small deconvolution kernel for image restoration has been sought based upon minimization of a target function, leading to a new restoration technique requiring considerably less computation compared to many other approaches. Regularization is achieved by introducing a multiplier that is in proportion to the average energy ofthe additive noise contained in the degraded image. The average noise energy may be estimated from the observed degraded image. Experimental results are given to show performance of the proposed method.

The disadvantage of the generalized learning vector quantization (GLVQ) and fuzzy generalization learning vector quantization (FGLVQ) algorithms is discussed. A revised GLVQ (RGLVQ) algorithm is proposed. Because the iterative coefficients of the proposed algorithms are properly bounded, the performance of our algorithms is invariant under uniform scaling of the entire data set unlike Pal's GLVQ, and the initial learning rate is not sensitive to the number of prototypes as Karayiannis's FGLVQ. The proposed algorithms are tested and evaluated using the iRIS data set. The efficiency of the proposed algorithms is also illustrated by their use in codebook design required for image compression based on vector quantization. The training time of RGLVQ algorithm is reduced by 20% as compared with Karayiannis's FGLVQ but the performance is similar.

For remote sensing imagery, every sensor system has unique system response, namely, point spread function (PSF), or modulation transfer function (MTF), which can be considered as sampling kernel given a prior. Sampling process like this doesn't satisfy Shonnon-Whittaker Representation Theorem's requirements, in such case, exact reconstruction is impossible. We have to look for optimal reconstruction in the sense of mean square, i.e. L 2 -norm. In this paper, we are mainly concerned with the applications of optimal reconstruction theory in remote sensing image processing, our aim is to develop a new resampling method for image magnification.

In fractal image compression, an image is coded as a set of contractive transformations, and is guaranteed to generate an approximation to the original image when iteration applied to any initial image. In this paper, according to Jacquin' 5 PIFS algorithm, and by analyzing traditional fractal mapping parameters, a kind of convolution-based fast fractal image coding scheme (CBFC) is advanced. To speed up the encoding and improve the compression ratio, it is combined with quad-tree partitioning neighbor searching algorithm. To improve the real-time performance of the algorithm, it is performed on TMS320C6201. Experiments results of algorithms based on CBFC, and CBFC using quad-tree partitioning structure on DSP are given in this paper as comparisons. The results show that fractal image real-time coding can be realized with the considerable reconstructive image and coding time.

The implementation of an image-based virtual navigation system is presented. An adaptive nonuniform sampling method of 4D light field is implemented based on slit images. Using the quadtree structure, this approach reduces the storage cost and accelerates the sampling process. The cause of holes is discussed and a hole-filling algorithm is designed. To avoid frame discontinuity, issues and approaches of memory management, track prediction and collision detection are discussed. Finally, an example navigation system of Science and Technology University of Macao shows the efficiency of the algorithms discussed.

In this paper, we propose an image compression method using the wavelet transform and context-based arithmetic coding exploiting bit plane conelation. The proposed method decomposes an image into several subband images using the discrete wavelet transform; the transform coefficients in each subband are classified into two classes; each subband image is then quantized; a "shuffling" process is applied to quantized wavelet coefficients; and finally arithmetic coding using the optimum context is carried out for bit plane coding of each subband. The performance improvement of the proposed method turns out to be mainly due to the "shuffling" process and the optimum context for arithmetic coding. Experimental results show that the proposed coder is similar to or superior to well-known existing coders, e.g. EZW and SPIHT, for images with low conelation.

Nowadays with the proliferation of readily available image content, image hiding has become a topic of very importance. In many applications, it is desirable to embed identifying images for authentication. This paper presents a novel digital image-hiding scheme based on blending technology in the frequency domain. The digital image blending is first introduced, and then the concept of digital image blending is generalized to the DCT domain. The properties of DCT blending are also studied. With these properties of the DCT blending, a novel digital image-hiding scheme is presented. The experimental results show that the scheme can be expediently realized and it is robust in a certain extent. The scheme also can be used in digital watermarking.

This paper presents a fast algorithm for fractal image compression. The algorithm uses quadtree partitioning to partition an image into image blocks of different sizes. Each of the image blocks is normalized to have zero mean and unity variance, and represented by a feature vector of dimension 1 6. The feature vectors, which can provide an accurate representation of the image blocks, are composed of the means and/or variances of each of the rows and columns. The k-d tree structure is used to partition the feature vectors of the domain blocks. This arrangement allows the search of the best matched domain block for a range block efficiently and accurately. An efficient encoding approach for low complexity range blocks is also proposed, which encodes the mean of a range block without searching the domain blocks. Moreover, during the range-domain matching process, a simple but very efficient search by using the property of zero contrast value is introduced, which can further improve the encoding time and compression ratio, especially in high compression ratio. This can lead to an improvement in encoding time and an increase in compression ratio, while maintaining comparable image quality. Experimental results show that the run-time required by our proposed algorithm is over 200 times faster than that of a full search.

Because of rapidness and easy to update, raster map is important in GIS and digital mapping. In many circumstances, the projection that users demand is not according with the original projection of the raster map, thus the projection transformation of raster map is necessary, Raster map projection transformation, which is essentially image transformation, involves transformation among four coordinate systems: original image coordinate system, original projection coordinate system, new projection coordinate system, new image coordinate system. In this paper, compared with the vast computation and the slow speed of traditional transform method, a rapid algorithm of raster map projection transformation based on dual transformation is researched. In this algorithm, first of all, transform rectangle control grid strictly according to the formula of projection transformation; and then, aiming at each rectangle, using the four corners to construct dual linear transformation polynomial between rectangle image and corresponding quadrilateral image in original raster map; finally, the direct transformation between image coordinates of two maps is realized. The experiment has proved that this algorithm not only satisfies precision demand, but also greatly improves transformation speed.

To effectively counteract the deceptions in multimedia ownership verification, the reversibility in the widely used and researched adaptive watermarking is investigated from a new point of view. The reversibility, together with its resulting attacks, is introduced and extended to adaptive systems. The mostly typical existing counteraction is revised and its potential is pointed out. Then, the intrinsic irreversibility of some adaptive technologies is disclosed and evaluated. Under these technologies, attackers have great difficulty dividing a released version into their claimed original data and scaled watermarks, and in the meantime making the latter be the adaptive results based on the former. Their reversed solutions are violently perturbed and perceptually unacceptable. The condition number of the coefficient matrix of the reverse equations can be employed to assess the degree of the perturbation. The experiments that use the proper adaptive filter to enlarge the condition number force the attackers to solve the difficult problems in algebra and signal processing. Instead of using specialized security processing, such as hashing and randomization, they exploit the irreversible nature of chosen technologies so that a more feasible irreversible watermarking scheme is achieved.

Network Keys Exchange Facility (NKEF) is a kind of negotiatory protocol. With it, network user can correspond with each other in ID authentication mode, encryption styles and secure connect time. It's one of research hotspots of network security problem. In this paper, Shamir protocol based scheme for Secret Transmission of digital image is proposed. According to the exchangeable character of encrypting operator, we give the scheme, which can overcome the traditional method's disadvantages including insecurity and that the amount of key is too large. This method takes full advantage ofShamir protocol' s skillful idea.

The basic concept of fractal and the mathematics of fractal compression are introduced in this paper, The basic principle and implementing method ofthe tradition fractal image compression are expound, A new method of sequence image fractal coding based on the visual character is proposed. In this coding algorithm, we use different fractal image coding method for different coding block after fully considering the statistic character, based on which we define the visual character of the coding block, then code the parameters of the iteration function by variably length coding. Consequently we obtain the higher compression ratio and the better decoding subjective quality.

VQ technology has been proved to be an important technology and are extensively used in the low-bit-rate image compression. The VQ quantizer consists of two procedures: an encoder and a decoder. The encoder assigns each input vector X to an index i, which points to the closest codeword Yi in the codebook. The decoder uses the index i to look up the codeword 1', in the codebook. With good designed coodbook, the VQ quantizer can obtain very low bit rate compressed image while the decoded images have high SNR. The VQ also leads to blocking effect which is not comfortable to naked eye. To smooth the decoded image, some image filters for color image are studied here. Linear filter, such as scalar/vector mean filter, can not fulfill the target. The nonlinear filter, such as scalar median filter, vector median filter and vector direction filter can do better work. The neural network image filter proposed in this paper has a forward structure. This filter is optimized with advanced GA method. Then this filter is applied to color image filter and have a better performance. The parallel structure makes it easy to implement on the DSP device. Both VQ encoder/decoder and NN image filter are implemented on the fastest DSP, TI TMS320C6416.

One of the most significant features of diagnostic echocardiographic images is to reduce speckle noise and make better image quality. In this paper we proposed a simple and effective filter design for image denoising and contrast enhancement based on multiscale wavelet denoising method. Wavelet threshold algorithms replace wavelet coefficients with small magnitude by zero and keep or shrink the other coefficients. This is basically a local procedure, since wavelet coefficients characterize the local regularity of a function. After we estimate distribution of noise within echocardiographic image, then apply to fitness Wavelet threshold algorithm. A common way of the estimating the speckle noise level in coherent imaging is to calculate the mean-to-standard-deviation ratio of the pixel intensity, often termed the Equivalent Number of Looks(ENL), over a uniform image area. Unfortunately, we found this measure not very robust mainly because of the difficulty to identify a uniform area in a real image. For this reason, we will only use here the S/MSE ratio and which corresponds to the standard SNR in case of additivie noise. We have simulated some echocardiographic images by specialized hardware for real-time applicati on;processing of a 512*512 images takes about 1 mm. Our experiments show that the optimal threshold level depends on the spectral content of the image. High spectral content tends to over-estimate the noise standard deviation estimation performed at the fmest level ofthe DWT. As a result, a lower threshold parameter is required to get the optimal S/MSE. The standard WCS theory predicts a threshold that depends on the number of signal samples only.

This paper introduces a two-phase algorithm to extract a center-adjusted, one-voxel-thick line representation of cerebral vascular trees from volume angiograms coded in gray-scale intensity. The first stage extracts and arranges the vessel system in the form of a directed graph whose nodes correspond to the cross sections of the vessels and whose node connectivity encodes their adjacency. The manual input reduces to the selection of two thresholds and the designation of a single initial point. In a second stage, each node is replaced by a centered voxel. The locations of the extracted centerlines are insensitive to noise and to the thresholds used. The overall computational cost is linear, of the order of the size of the input image. An example is provided which demonstrates the result of the algorithm applied to actual data. While being developed to reconstruct a line representation of a vessel network, the proposed algorithm can also be used to estimate quantitative features in any 2-D and/or 3-D intensity images. This technique is sufficiently fast to process large 3-D images at interactive rates using commodity computers.

Three kinds of image reconstruct algorithms for Electrical Resistance Tomography (ERT) has been researched, and a new ERT reconstruct algorithm-Regularized general inverse(RGI) ERT reconstruct algorithm is proposed, which is based on linearity ERT forward problem, and makes use of general inverse to confirm the minimum norm error solution of ERT inverse problem. Meanwhile, adopting regularized method to stabilized the numerical value. The observation operator is set up by multiple linear regression method. Three restriction conditions is brought to bear the optimum stabilization solution. The simulation result shows that reconstructed image can reflect the truth medium distribution in the field truly including different complex distributions. After filtering the images by unite bound for the same medium distribution, the average of CSIE image reconstructed by linear back project algorithm, sensitivity coefficient algorithm and regularized general inverse algorithm is 12%, 9% and 6% respectively. The result shows that the image quality reconstructed by regularized general inverse algorithm is improved in evidence than that of the other two algorithms. The calculate amount of regularized general inverse algorithm is same as one step sensitivity coefficient algorithm, the speed of reconstruction is fast.

In this paper, we present a novel digital watermark system frameworkfor 3D mesh model with all mesh attributes, including position coordinates, normal, color and texture coordinates etc. Within this framework, 3D mesh model attributes are considered as geometry signals defined on mesh surfaces. A planar parameterization algorithm, which is first proposed by us, is used to map 3D mesh models to 2D parametric meshes. Geometric signals are then transformed into 2D signals. Then a wavelet-based watermark casting scheme is proposed to embed the watermark into some wavelet coefficients. Experimental results show that the embedded watermark is robust against various geometry signal processing.

Thresholding is difficult for images with poor contrast or illumination, intensive noise and non-planar background. An active surface based adaptive thresholding algorithm is proposed in this paper. Derived from the idea of active contour models, an active surface model is used to estimate the background surface ofthe image. Subtraction ofthis active surface from the original image surface is to remove the influence of uneven background and poor illumination, and convert the problem to a global threshold one. Thus a proper choice of the global threshold will obtain a desirable binary result.

By replacing the complex-valued Gabor basis functions of the complex-valued discrete Gabor transforms (CDGTs) with real-valued Gabor basis functions, we propose fast algorithms for 1 -D and 2-D real-valued discrete Gabor transforms (RDGTs) in this paper. The RDGT algorithms provide a simpler method than the CDGT algorithms to calculate the transform (or Gabor) coefficients of a signal or an image from finite summations and to reconstruct the original signal or image exactly from the computed transform coefficients. The similarity between the RDGTs and the discrete Hartley transforms (DHTs) enables the RDGTs to utilize the fast DHT algorithms for fast computation. Moreover, the RDGTs have a simple relationship with the CDGTs such that the CDGT coefficients can be directly computed from the RDGT coefficients.

Codeword index assignment (CIA) is a key issue to vector quantization (VQ). The application of tabu search algorithm has made some achievement in solving the codeword index assignment. Combined with the notion of tabu search the energy allocation scheme has been also successfully used to overcome channel error sensitivity . In this paper, a new algorithm called modified tabu energy allocation algorithm (MTEAA) is applied to index transmission of codeword for the purpose ofminimizing the extra distortion due to bit errors. Simulated annealing (SA) technique and a new parameter are introduced in the iteration of the tabu energy allocation approach (TEAA) to improve the convergence performance. Experimental results demonstrate that the proposed algorithm is superior to TEAA and the conventional energy allocation algorithm (CEAA) by evaluating the performance of channel distortion.

The automatic sorting of bank bill has become very important to speed up the office automation. Some methods have been reported to classify the denomination and directions of bank bill. But there were no reports about defile detection. In this paper, we designed a fast and effective algorithm to detect defiles on bank bill. We make use of mathematical morphology image processing to do shift and rotation correction. In order to reduce CPU time, we adopt pyramidal strategy to do template matching. In addition to these, we designed a check-up means to eliminate dummy defiles. The test results prove that this algorithm is robust. It's not sensitive to random noise and geometry distortion. In a word, this algorithm is quite effective.

Based on evolutionary computation, a new classification approach, OCEC, for radar target recognition using high-resolution range profiles is proposed. OCEC has nothing with dimension ofthe input space and has advantages of small amount of computation and stable performance. Simulations are presented to classify microwave anechoic chamber data for three different scaled airplane models. The results show OCEC has higher performance than other algorithms for recognition ofhigh-resolution radar target range profiles.

A simple and effective error detection and error concealment based on the H.263 video decoder has been described in this paper. The H.263 syntax structure and semantics, VLC code word and continuity property of video signal at the video decoder are employed to process the error detection. The spatial and temporal correction is made full of use for error concealment. It has been found that this method can greatly improve the quality of the reconstruction images without decreasing the coding efficiency.

This paper proposes an optimization model for extracting edges in gray-scale images. The model sufficiently utilizes the gray-level information in a pair of orthogonal directions at each considered pixel. The model has three major features in its novelty: (1) Emphasizing the globality of traditional local features; (2) Being a generalized case of the classical snake models; and (3) Offering a theoretical interpretation to the setting of the parameters for the method based on the Simulation of Particle Motion in a Vector image Field (SPMVIF). Our Edge Detection from Edge Knots (EDEK) model can be divided into two stages: the generation ofedge knots on or near edges and a propagation process for producing complete edges from these edge knots. One advantage ofour approach is that the propagation process does not depend on any control parameters. The EDEK model is suitable for live plant image processing, which is demonstrated by a number of simulation results on the edge detection of live plant images. Our model is simple in computing and robust, and can perform very well even in situations where high curvature exists.

This paper studies restoration of images blurred caused by uniform linear motion and its spatial domain processing. Horizontal motion is taken as the canon case, 1 -dimensional traveling wave equation is adopted as the mathematical model for image blurring, and Hough transformation is applied to detection of the motion direction. As for oblique motions, they are transferred to horizontal ones for easy solution. Experimental results show that the deblurred images resulting from this spatial domain processing are of better quality than that from traditional frequency domain processing where the ring effects will more or less appear and the computing cost is high.

Landsat TM image is the most popular and universal RS information source, and got wide uses in different fields such as resource investigating, environment monitoring, urban planning, disaster preventing and as on. Although TM image has got wide applications, its use in mining area is still in experiment and beginning stage because mining area is a kind of special and complex geographic region. One of the most important issues is to study the information charaéteristics and determine the most effective band combinations oriented to given region and task. In this paper, Xuzhou mining area, located in Northern Jiangsu Province, is taken as the studying area, and Terrestrial Surface Evolution (TSE) as the studying task. According to the specific condition of studying area, the information characteristics of each band of TM image and relations between different bands are analyzed by selecting different sampling area, and relative rules are given. After that, band combination is discussed and the information content is used as the judging rule. Because more bands will require more computer resource and is low speed and cost consuming, three-band combination is used widely. It is found that in all three-band combination schemes, the combination of Band 3, Band 4 and Band 5 is the most effective. Finally, Genetic Algorithm (GA) is used to the band selection in multi-band RS image, and it proved that GA is an effective method to determine the optimal band combination, especially for multi-spectrum and super-spectrum RS information source, and GA is also a good optimized algorithm in Geoscience.

The local accumulative histogram gives us the statistic characteristics about the image more flexibly and precisely than the whole histogram of the image. We use the intrinsic structure of the edge region alternated with steep rising and stable rising to detect the edge information, which is the key feature in the local accumulative histogram. Then we refine the edge image to one-pixel wide edge image by the threshold obtained from the local accumulative histogram. The experimental results show that not only the algorithm can give us a good edge detecting result, but also it has the ability of antinoise. And some definitions of the one-pixel wide edge are given out in accordance with the people's intuition more directly and simply than other definitions before.

In this paper, we propose a multi-scale image segmentation algorithm based on mean shift, a simple iterative procedure that shifts each data point to the average of data points in its neighborhood, which has been proven to be a mode-seeking process on a surface constructed with a "shadow" kernel. In the presented algorithm, not only the color features, but also the space relationship of each pixel are considered in multiple scales, thus getting a more reasonable clustering sequence, furthermore, center candidates are validated by contour map. Experimental examples are illustrated and compared to show that the approach is effective not only in segmentation, but also in denoising.

Quality assurance is the key for increasing competition in the market place. This paper presents a new machine vision based approach for the detection of defects using real Gabor functions. A bank of real Gabor functions, followed by a nonlinear function, is used to sample texture features at different scales. These texture features are compared with those from defect-free (reference) image, and a set of feature difference arrays are created. These are used to generate a combined image output using image fusion. This combined image output is used to obtain a binary image of defects using calibration. This paper also details a new method for automated selection of the center frequency of Gabor function using spectral analysis. Experimental results have confirmed the usefulness of the proposed approach for the automated inspection of textile webs.

There are different techniques available for solving of the restoration problem including Fourier domain techniques, regularization methods, recursive and iterative filters to name a few. But without knowing at least approximate parameters of the blur, these methods often show poor results. If incorrect blur model is chosen then the image will be rather distorted much more than restored. The original solution of the blur and blur parameters identification problem is presented in this paper. A neural network based on multi-valued neurons is used for the blur and blur parameters identification. It is shown that it is possible to identify the type of the distorting operator by using simple single-layered neural network. Four types of blur operators are considered: defocus, rectangular, motion, and Gaussian ones. The parameters of the corresponding operator are identified by using a similar neural network. After identification of the blur type and its parameters the image can be restored using different methods. Some fundamentals of image restoration techniques are also considered.

In this paper, we present a multi-scale internal energy model for the balloons, which are closed snakes driven by pressure forces. When used to extract fuzzy contours from noisy images, a balloon sometimes goes through gaps on the contour, thus not able to reach equilibrium. We argue that a balloon should be discretized with a scale comparable to sizes of gaps for the internal energy to be effective in performing contour completion. However, using too large a scale will also prevent a balloon from catching necessary details. To solve this problem, we propose a multi-scale internal energy model that has the ability to maintain smoothness at various scales, thus able to perform contour completion for gaps with various sizes without missing any details.

One of the most popular level set algorithms is the so-called fast marching method. In this paper, a medical image segmentation method is proposed based on the combination of fast marching method and watershed transformation. First, the original image is smoothed using nonlinear diffusion filter, then the smoothed image is over-segmented by the watershed algorithm. Last, the image is segmented automatically using the modified fast marching method. Due to introducing over-segmentation, the arrival time the seeded point to the boundary of region should be calculated. For other pixels inside the region of the seeded point, the arrival time is not calculated because of the region homogeneity. So the algorithm's speed improves greatly. Moreover, the speed function is defmed again based on the statistical similarity degree of the nearby regions. Experiments show that the algorithm can fast and accurately obtain segmentation results ofmedical images.

Based on radar remote sensing and sediment facies analyzing, this paper studies the features of environmental evolution of North-eastern Ejin Banner since 60 Ka BP(Before Present). The conclusions are listed as follows: (1) The evolution of Gaxunnur Lake, Sugunur Lake, Tiane Lake is dominated by faults and regional climate; (2) By analyzing deposit of old Juyanze Lake, it is a large outlet lake about 5OKa BP. Just about 5OKa BP, there was a rapid decline of temperature in the Northwestern of China. The events caused those lakes shrinkage. (3) By fault activity arising uplift in the north of old Juyan Lake and depression in the south, the lake's water flowed out from north to south at around 35Ka BP, which is reversed to the former flow direction. So, old Juyanze fluvial fan was formed. At the same time, Juyan Lake separated from Sugunur lake and Wentugunr old channel was abandoned. (4) Near 2000 years , Ruoshui River is a wandering river. Due to intense influence of human activities, the oasis ecosystem is rapidly degenerated in recently 15 years.

In the close digital photogrammetric three-dimension coordinates measurement, the circular target is often taken as imaging feature and mounted on the measured object or the probe for 3D coordinates detection. The accuracy with which circular targets are located determines the effectiveness of measurement. Subpixel level accuracy is one of the methods that can improve the accuracy of target location. Many methods which based on subpixel edge or centroid detection have been developed and analyzed for target location, but little research focused on circular optical target location. In this research, a new algorithm named bilinear interpolation centroid algorithm was developed for circular optical target subpixel location. In this technique, the accuracy of the squared gray weighted centroid algorithm can be improved by increasing the available pixels which obtained by bilinear interpolation. The intensity profile of the imaging points and the signal to noise ratio, which affect the subpixel location accuracy, are optimized by automatic exposure control. The experiments have shown that the accuracy of this algorithm is better than the traditional centroid algorithm, and the absolute error of less than 0.0 1 pixels is obtained on the image of a rigid reference bar.

s+P transform proposed by Amir Said[7] is a multiresolution representation method. It can map integers to integers, so it has been used in lossless compression of images with better performance than that of JPEG standard based on linear prediction. To further improve compression performance, this paper presents an efficient lossless compression algorithm of images based on DPCM and S+P transform. Firstly, an error image is obtained from an original image with linear prediction; Secondly, the error image is transformed with S+P transform; Finally, the transformed coefficients will be compressed effectively with entropy coding. We list our software simulation results, which are compared with those of other known algorithms7111011121 for multiresolution-based lossless compression. The comparison shows that our new method is effective and efficient, and produces the best results.

The application of image morphing to computer animation and computer graphics is experiencing broad growth. It has proven to be a powerful visual effects tool in film and television, depicting the fluid transformation of one digital image into another. In this paper, the authors present a new method for image morphing based on field morphing and mesh warping. Our method makes use of the feature specification method of field morphing which is simple and expressive and the warp generation approach of mesh warping which is straightforward and fast. Some measures are taken to make the proposed method work; experimental results show that the proposed method facilitates the input of features and calculates quickly with a steady metamorphosis effect.

We present a joint segmentation scheme for the stained renal tubular image in this paper. The scheme, which combines several distinct image processing methods and uses a special criteria to select appropriate segmented areas based on regional properties, makes the best use of different segmentation methods. Experimental results are given and some further improvements are discussed at the end of this paper.

In this paper a framework of a wireless spread spectrum video system has been proposed. Because the energy supply in the wireless systems is limited, the power consumption of the systems must be reduced. Consequently the design principles for low-power wireless video systems are introduced at first. Wireless video systems have many requirements on encoding/decoding. And MPEG-4 standard is just suitable for the systems by reason of its high compression efficiency and outstanding performance. Thus the video information can be transmitted at a very low bit rate over the wireless channels. Given the advantages of the spread spectrum system, we also introduced spread spectrum into the wireless video systems and gave a framework of wireless spread spectrum video system.

This paper presents a private and lossless digital image watermarking system. First, to ensure the security of the watermark, the watermark image is scrambled before embedding. And then, Human Visual System (HVS) is employed to further enhance the transparence and the robustness of the watermarking system. Thirdly, the watermark image is compressed as watermarking information and is embedded in the DCT domain of the original image. Finally, We save the decimal fractions of the image-signed as a data file as a key during embedding and the data file can avoid the decimal fractions data loss caused by saving the image-signed. The experimental results show that it can embed the image as the watermark with strong robustness.

in the paper a theoretical analysis ofoptimal cubic filter for image scaling is investigated. Starting from the analysis of the cubic filter in the frequency domain, and then by minimizing the magnitude of the frequency response, |H(?)|, over the interval 0.5 ??? 1.0, the best choice of the coefficients for the cubic filter can be obtained in terms of aliasing suppression and clear text display. By incorporating the previous result obtained by Keys with the new result obtained here, the optimal coefficients for the cubic filter can be obtained. Since the optimal cubic filter is obtained by suppressing the frequency response beyond the Nyquist frequency, it can be expected that the aliasing can be suppressed by using the proposed optimal cubic filter. The proposed optimal filter has performed a text enlargement example, and simulation results justify our theoretical analysis. it shows that the proposed method has very good performance in image scaling, especially in performing text enhancement.

The purpose ofthis note is to present a novel techniquefor boundary determination that incorporates both a shape prior and a local intensity profile into a geometric active contour model. The basic idea of this model is to minimize an energyfunctional, which incorporates the information ofthe image gradient, the shape ofinterest, and the intensity profile across the shape into the search for the boundary of an object. The level set form of the proposed model is also provided. Experimental results on ultrasound images are presented, which indicate that this model provides close agreement with expert traced borders. These results indicate the effectiveness ofthe incorporation ofthe local intensity profile in the model.

Five robot vision algorithms for object segmentation are explained. This algorithms are compared in an experiment. The aim of the experiment was to check which of this five algorithms are useful for our robotics project in the field of indoor room exploration. In this project a robot shall autonomously generate a map from our office which shall be the foundation to fulfill tasks like an office messenger. Two evaluation criterions were used to get hints with regard to the eligibility if inhomogeneous illumination must be handled. Although the number of researchers, which are working on the field of video based exploration with autonomous robots, is increasing, it don't exist sufficient works at this juncture which give hints how an appropriate evaluation ofrobot vision programs in autonomous robot systems can happen. This work wants to give contribution to this gap.

This paper discusses about ergodic matrices and related application about scrambling and encryption of digital images. First, we use ergodic matrices to realize the position permutation algorithms. In particular several novel methods of scrambling are proposed. Subsequently we analyze the isomorphism relationship between ergodic matrices and permutation group. By defining permutation symbol operation system of ergodic matrices, we construct a union form to express all existed permutation algorithms. Finally the images are encrypted using them. Preliminary results are satisfactory.

An improved multiscale image enhancement algorithm based on Laplacian pyramid (LP) is described. At each scale of the LP, the local variance threshold and relative enhancement are implemented by modifying the detail coefficients of the LP nonlinearly. With the local variance threshold, contrast enhancement occurs in high detail areas and little or no image sharpening occurs in smooth areas. And the objective of relative enhancement is to enhance the details with lower magnitude more than the details with higher magnitude in each scale of the LP. Since the low scales of the LP have subtler image features, we modify the local variance threshold and relative enhancement to take the different significance of different scale into account. So the low scales are more enhanced than the high ones. The given enhancement algorithm is simple to implement and suitable for generic image including CT and X-ray images. Experimental results show that the contrast improvement ratios of most images are increased while preserving the good visual assessment of original image.

The estimation of cloud motion from a sequence of satellite images can be considered a challenging task due to the complexity of phenomena implied. Being a non-rigid motion and implying non-linear events, most motion models are not suitable and new algorithms have to be developed. We propose a novel technique, combining a Block Matching Algorithm (BMA) and a best candidate block search along with a vector median regularisation.

This paper discusses the restoration of blurred images caused by arbitrary direction constant-speed straight-line movement. Estimating the PSF of a real-world motion-blurred image is an essential step in the restoration process. In the paper, we first prove that we must set two-dimensional point spread function and directly restore blurred image in 2D form, instead of first restore in X direction and then in Y direction. Then we propose the spectral method for estimating the blur range BE and direction of PSE. The DFT image of blurred image contains parallel strips vertical to motion direction. The distance of G(u, v) image center to neighboring black line is reversedly proportional to blur extent BE. Finally, the validity of our PSF estimating method and superiority of windowed Wiener filter over standard Wiener filter is clearly demonstrated in section 4.

This paper presents the auto-matching algorithm and applies it to human face warping. In this algorithm, we first prewarp the two origin images to make their epipolar lines parallel to the horizontal scan-lines. Then we detect edge points on the scan-lines of the first image by the use of the Gauss first differential filter and match the corresponding points on the same scan-lines of the other image. With these corresponding edge points, we can segment the scan-lines of the two images. In the end, we use affme transformation to obtain more exact corresponding relation of these points. Some experiments have proved that the algorithm is simple and doable.

SAR (Synthetic Aperture Radar) and the related techniques for application have experienced great development. And much attention has been paid to speckle, which results from the coherent imaging process. Because speckle acts as a noise multiplying the underlying RCS (Radar Cross-Section), and it greatly degrades interpretation, identification and other quantitative application of SAR images. Diversified speckle reduction techniques have been proposed to reconstruct the mderlying RCS from the observed SAR image. So the need for assessment criteria of speckle filters becomes pressing. Based on study of properties of SAR imagery and the existent filters, evaluation criteria of speckle filters are analyzed, and quantitative indexes can be adopted in practical application are proposed according to the criteria. Radiometric and spatial indexes for quality of SAR images are applied to speckle reduction in uniform areas and isolated scatterer preservation. Reasonable indexes are defined for edge preservation and radiometric distortion, respectively. Simulated and real SAR images are used for experiment. The experimental results accord with theoretical analysis of the testing filters. This agreement demonstrates the effectiveness f the proposed indexes. Distribution of speckle image suggests another way for evaluation, but further study is needed to bring it to practical application by ?2 goodness-fit-test.

Digital image information security is a new research topic which is appearing with the rapid progress of computer network and multimedia. Becoming main measure of information representation, digital image meets the problem how to protect the security of the information it carries. In particular, how to keep the information after kinds of image processing procedure is still an open problem which is not solved. The common way to hide information into an image is to change the values of certain places in the space domain or frequency domain. The novel technique introduced in this thesis, also being a new idea, is to use the topological characteristics of certain chosen points, in detail, the sequence of points along a convex-chain, in the picture to carry information. And this makes the information quite hard to be destroyed by many kinds of image processing.

Fuzzy C ellipsoidal (spherical) shell cluster algorithms are established for clustering shell shape patterns. In an application to blood cell image processing, by several steps of proper processing of blood cell images using morphological operations, we thus obtain a prior knowledge for the cluster prototypes, and training sets with spherical shell shape. The number and radiuses d blood cells can be calculated with a few iterations using introduced approach.

A new method is proposed when optical flow is estimated in a color image sequence. That is the indirect method exploiting Hue and Saturation as color invariance under motion, with differs from the straightforward method of using the color components as separate images ofthe same scene. Also the two color systems HSV and HLS are used for two reasons. Firstly, they have advantage over the traditional method in using the ratio of the color components as separate images of the same scene, so that they can avoid linearly dependent of color components. Secondly, they correspond to human vision since the two color systems have been developed for user and can represent the color characteristics. In order to prove the new method effective, computer simulation experiments were carried out while estimated the optical flow in color synthetic sequences and analysis their performance of two -frame color optical flow both qualitatively and quantitatively. The results showed that the new method yields accurate optical flow field estimation.

A two-step non-linear medical image registration approach is proposed, based on the image intensity. In the first step, the global affme medical image registration is used to establish one-to-one mapping between the two images to be registered. After this first step, the images are registered up to small local elastic deformation. Then the mapped images are used as inputs in the second step, during which, the study image is modeled as elastic sheet by being divided into several sub-images. Moving the individual sub-image in the reference image, the local displacement vectors are found and the global elastic transformation is achieved by assimilating all of the local transformation into a continuous transformation. This algorithm has been tested by both simulated and clinical tomographic images.

This paper uses semivariogram to quantify the crop spatial pattern from ERS-2 SAR image, especially for the cotton field, to improve the extraction accuracy for cotton growth monitoring. Measuring the influence of the semivariogram calculation variable can understand and control the calculation variable for remote sensing classification better. The crops semivariograms of study area exhibit a similar bounded shape resulting the regularization effect, the sill reaches at about 12 pixel, 150 m, the mean size of agricultural fields in the studied area. In this agricultural landscape, spatial structure results mainly from cultivation patterns. The cotton and maize semivariograms are quite different distinctively. The semivariogram of each class reflects the texture characters, it measures the each class spatial structure and similarity relative to the size and direction of calculation window, which has different effect on the results of classification. We can select the window size according to the range of each class. Joining the classification with the average value for the four direction semivariograms can reduce the band numbers and classification time and elevate the accuracy. The results in study area indicate combining average semivariogram and spectrum in classification elevates 12.4% on overall accuracy compared to spectrum only.

Independent Component Analysis (ICA) is a new signal processing method developed recently which analyzes the data from a statistical point of view. In ICA, one can try to express a set of random variables as linear combinations of statistically independent components. In this paper, ICA is applied to image feature extraction, and the information maximization algorithm is performed to optimize the results. From the results, it can be seen that the extracted features represent the image data in a natural way. In addition, the ICA basis vectors are localized and oriented, and sensitive to lines and edges of varying thickness of images. As an application of these extracted features, another denoising experiment is done. In this experiment a Gaussian noise is reduced by applying a soft-thresholding operator on the extracted ICA coefficients.

MPEG-4 provides a basic tool for interactivity and manipulation of video sequences. To take advantage of these content-based functionalities, video sequences must be segmented into semantically meaningful objects. Video object segmentation is a key step in defining the content of any video sequences. The algorithm proposed in this paper is a spatiotemporal segmentation. It starts from an over-segmented image by morphological gradients, and then the segments are merged by spatiotemporal information. To tracking the segmented objects, stochastic optimization methods are used to form homogeneous dense optical vector fields. We simulate the algorithm in the Cellular Neural Networks (CNNs) architecture by MATCNN. It suggests a fully parallel implementation in CNN-UM chip.

In this paper, we present an image segmentation approach, which is based on the data fusion technique and divided into two steps. At first, we segment the image in a low-resolution to find the coarse contour of regions; then we use the fuzzy-c-means algorithm to process the image in a fine resolution to find its delicate characters. To avoid the large amount of calculation, the results of the first step is fused into the second step. Experimental results show that this technique is effective in improving the quality of segmentation and lessening the calculating time.

In this paper, an image quality measure, termed pixel-based correlation weighted-mean square error (WMSE), is presented. The proposed distortion measure depends not only on the mean square error in the distorted image, but also on correlation between pixels in a predefined neighborhood. Experimental results are given to show the advantage of the described method.

In this paper, a new algorithm based on the inherent characteristics of specific objects is proposed, which can solve the rotation angel and scaling between a Real Aperture Radar (RAR) image and a reference image. This approach can also significantly alleviate the succeeding calculation duty of matching. Experiment shows that the algorithm developed is reliable and practical.

A simulated annealing based coefficient optimization approach to improve the image fusion performance is proposed in this article. This article tries to answer two questions. Firstly, reference images used in most previous literature to measure the performance of fusion process are often real scene image or the equivalence, which hardly exists in practice. To avoid using the nonexistent real scene image, an alternative reference is created through extracting and merging edge features from each input image. Secondly, the performance measurement needs to be involved in optimization process. Our method adopts a simulated annealing procedure to optimize the Wavelet coefficients combination. As the objective function, the cumulative edge distance is minimized by adjusting the magnitude of the wavelet coefficients according to the error matrix, so as to approaching the optimal coefficient combination.

Principles of colour science have been widely used in applications of remote sensing, but few papers have reported the quantitative application of color principles in the processing of multi-spectral remote sensing images. in this paper, authors have developed a colour difference technique for the classification of multi-spectral remote sensing image according to the principles of colour science. it is demonstrated that the colour difference technique can be used in the classification of IKONOS multi-spectral remote sensing image, the results of image classification are correspond to human eye 's interpretation very well. in addition, the colour difference classification technique is very sensible to chromatic information in the multi-spectral remote sensing image, but has a luminance latitude. Therefore, the colour difference classification technique seems very adaptable to the classification ofuneven irradiance multi-spectral remote sensing image.

There are abundant metallic mineral resources on the sea floor, polymetallic nodule is an important one of them. A polymetallic nodule ore includes nickel, copper and manganese elements, etc. So the polymetallic nodule is very important and precious to industry. In order to know the distribution and reserves in the west and east pacific areas, a deep-tow optic system is imported from U.S. to acquire deep sea-floor images. Processing the images, we can extract some information and calculate some parameters: coverage, grain size and abundance, which stand for distribution and reserves of the polymetallic nodule. In the paper, features of the deep sea-floor image are analyzed, considering the characters, a processing procedure for the deep sea-floor pictures is presented. Methods are presented to rectify radiative uneven and geometric distortions, at last, the correlations of coverage, abundance and grain size are analyzed and the formulas for computing abundance are respectively derived from that.

Computer Assisted Chromosome Analysis can effectively alleviate the doctor's burden and keep the analysis accuracy. In view ofcurrent situation in China and abroad, a new microcomputer-based chromosome analysis system is presented in this paper, which is simultaneously cheap, convenient, and effective. At the same time, its automatization and analysis capability are higher than other system's in China, as a result doctors can use it to analyze chromosomes easily and accurately.

Image segmentation is the pre-step of multi-target tracing in Computer Assisted Sperm Motion Analysis System (CASMA). As a special sperm-tracing problem, a fast, automatic, unsupervised segmentation algorithm is required. In this paper, we utilize four segmentation algorithms to segment three different kinds of sperm images sampled from our actual system. By making an overall comparison between them, a conclusion is reached that the Otsu's maximum between-class variance algorithm is the most suitable for the special sperm microscopic image segmentation and this segmentation algorithm has been successfully applied to our developed system.

The definition of scene navigability is set force, and acquisition probability is regarded as an important mark of scene navigability. The preprocessed reference map and additional noise are supposed as independent random field with zero mean. An estimation of acquisition probability on Cross Correlation Scene Matching Systems is derived by analyzing context and statistics characteristic of reference image. Many experiments with satellite images demonstrate the validity of this proposed acquisition probability model, which provides a theoretical and practical basis for design of self-guided systems using scene matching.

Extracting features from fingerprints is a crucial step in fingerprint verification and recognition. Here, features are represented as minutiae, such as end points and bifurcation points. Many algorithms for this issue have been developed recently. This paper presents a fingerprint feature extraction method through which minutiae are extracted directly from original gray-level fingerprint images without binarization and thinning. Our algorithm improves the performance of the existing ones along this stream. Our experimental results demonstrate that our approach can achieve better performance in both efficiency and robustness.

In this correspondence, a new block sum pyramid algorithm (NBPSA) to motion estimation is presented. Compared with BSPA, NBSPA estimate the vector of the minimum mean absolute difference ( MADmin ) In the mean time, instead of update the value level by level, we update the estimation of MAD row by row, up to down. Experimental results showed that, with the search result ofthe algorithm identical to the search result ofthe exhaustive search and BSPA, NBSPA reduced the computation complexity greatly.

Microarray is a widely used method in molecular biology. The essential problems of microarray image analysis are to locate spots and quantify the intensity of each spots. Because of the rapid progress in gene technology, fully automated microarray image analysis is required to process thousands upon thousands images produced in hybridization experiments. In this paper we address an efficient and fttlly automated method to solve this problem. The only precondition of our method is that the spots are arrayed in a grid. An implemented program based on our method can cope with the following problems: global rotation of the grid, absence of grid spots, and local deviation of the spot from its specified grid position. In grid fitting step, we imported a method introduced in document layout analysis system to estimate the rotation angle and the grid unit size, spots location amplification is then performed to locate all the grid elements. And in quantification, we use the local threshold method to make the simplest integration method getting a reliable result. All algorithms used in the program can efficiently run without operator' s intervention.

An advanced Independent Component Analysis (ICA) algorithm based on genetic algorithm is proposed with analysis to the ICA method. The proposed algorithm can be used to solve the problem of local optimum that is easily stacked into by normal numerical solution. The image separation simulation shows that the global optimum can be acquired through the proposed algorithm under the circumstance of adequate colony size and genetic generations.

Paimprint is a new biometric method to recognize a person. The most important feature of paimprint is the lines. In this paper, a set of line detector is devised for paimprint. There are two parameters in these detectors, one controls the smoothness and connection of the lines, the other controls the width of lines which can be detected. The lines in different directions are detected by corresponding direction detectors and then fused into one edge image. In training stage, the lines of the training samples are represented and stored with chain code. In the verification stage, the lines are matched using Hausdorif distance. Experimental results show the efficiency of this method.

In this paper we present a fast iterative image superresolution algorithm using preconditioned conjugate gradient method. To avoid explicitly computing the tolerance in the inverse filter based preconditioner scheme,1 a new Wiener filter based preconditioner for the conjugate gradient method is proposed to speed up the convergence. The circulant-block structure of the preconditioner allows efficient implementation using Fast Fourier Transform. Effectiveness of the preconditioner is demonstrated by superresolution results for simulated image sequences.

Thinning algorithms are widely used as a useful method of pre-processing in image process. In this paper, a high-speed thinning algorithm for character recognition is proposed. This method can be used by any thinning algorithms for speeding process without any changes of the thinning algorithms. The Hilditch thinning algorithm is adopted to verify the proposed method. With a 3x3 maskimage in the Hilditch thinning algorithm, the result ofprocess output can be saved to a memory 'table. The output results of all different 3x3 masks are saved to this 'table" at the beginning of character recognition. When an image will be processed, the thinning results of every 3 x 3 masks in the image can be extracted by the method of "looking for table". Thus the thinning result is same but the process speed is high. The proposed method has been evaluated by comparing the time cost of the proposed method and Hilditch thinning algorithm using different configured computers to process a same image and the result is satisfied.

This paper presents a fuzzy information fusion method to automatically extract tumor areas of human brain from multispectral magnetic resonance (MR) images. The multispectral images consist of T1 -weighted (T1), proton density (PD), and 12-weighted (T2) feature images, in which signal intensities of a tumor are different. Some tissue is more visible in one image type than the others. The fusion of information is therefore necessary. Our method, based on the fusion of information, model the fuzzy information about the tumor by membership functions. Thismodelisation is based on the a priori knowledge of radiology experts and the MR signals of the brain tissues. Three membership functions related to the three images types are proposed according to their characteristics. The brain extraction is then carried out by using the fusion of all three fuzzy information. The experimental results (based on 5 patients studied) show a mean false-negative of 2% and a mean false-positive of 1 .3%, comparing to the results obtained by a radiology using manual tracing.

An effective immune cell image segmentation algorithm based on mathematical morphology is presented in this paper. In order to get better segmentation results in addition to the morphology based watershed growth algorithm the histogram potential function is involved, which means, the image spectral information is combined with spacial information. How to get the exact segmentation result is a major issue for immune cell image analysis. Obtaining an effective and credible marker is a crucial step of watershed segmentation. By involving the histogram potential function, the markers suitable for watershed segmentation can be clearly improved and the segmentation result is quite consistent with human vision and also the segmentation speed and repeatability are quite acceptable.

An intelligent image-indexing algorithm is proposed in this paper. It based on knowledge extracted from some simple single low-level image features. Two independent large image databases are built with more than 12000 images for training and test, and the experimental results show it work efficiently for high dimension database indexing. The running time is shorter than other algorithms proposed for the same purpose, and the algorithm performs even better for some certain semantic image classifications.

Hough transform is a traditional algorithm for line detection. But this algorithm has many disadvantages, such as large computing load, long operating time, large occupied memory and low precision. This paper forwards a new line detection algorithm, that is, detecting line based on random sample theory. This algorithm can solve the problems brought by Hough transform. In this paper, at first, the steps ofboth Hough transform algorithm and random sample line detection algorithm are listed respectively and the latter one is discussed in detailed. Then, the two algorithms are tested by the simulative experiments. The results of these experiments are compared and analyzed. At last, the result of the experiment using real image is displayed.

We present a method to locate the caption area of frames in videos. Histogram can be extracted easily from an image, while the color of caption area is apparently brighter than other areas. Unlike previously published methods of using edge detection methods to detect the edge of caption from images, this method uses the caption histogram to detect the caption area, and applies the probabilistic theory to determine the area based on the caption histogram. Experimental results show that our method is practicable and prosperous.

Aiming at the existing questions in the preprocessing step of fingerprint image, an improved edge detection algorithm is presented based on the concept of fractal Brownian motion model in this paper. Due to the fact that the structure of fingerprint is typically self-similar, which satisfies the model of fractional Brownian random field except edges, so it can be used as the principle of detection. By selecting those pixels that fractal parameters are in a certain range, the fingerprint edges can be detected. The results of experiment have shown that this improved algorithm obtain much more edge details than most traditional methods do and reflect the edge information of original image much better. At the same time, this algorithm reduces the computation complexity greatly and enhances the speed of edge detection.

Clustered microcalcifications (MCCs) on mammograms are important hints of breast cancer. Nevertheless, it is a complex and difficult task for radiologists to detect the clustered MCCs from the tissue background of mammograms only by naked eyes. This paper presents a method for computer-aided detection of MCCs in digital mammograms. The detection algorithm mainly consists of two different methods. The first one, based on the difference-image technique, recognizes high-frequency signals and very high-frequency noise. The second one is able to extract high-frequency signal by exploiting a wavelet based noise suppression and neural network (NN) classification. In the false-positive reduction step, false signals are separated from MCCs by means of an AND operation on signals from two methods. The algorithm is tested with a series of clinical mammograms. A sensitivity ofmore than 90% is obtained at a relatively low false-positive (FP) detection of 2.18 per image. The results are compared with thejudgement ofradiological experts, and they are very encouraging.

In this paper, a normalized RGB space based color clustering method is given, and further the application of color clustering in character location is described. Color clustering is used to group a color image into the different binary layers. During color clustering, the normalized RGB space is adopted. In each layer, its color should be homogeneous. As characters normally have different information in color from their background, characters and their background are grouped into different color layers, which are fairly useful to locate characters. In order to achieve character location, an aligning and merging analysis (AMA) scheme is presented to locate all potential characters on each color layer. The experimental results have proven the effectiveness of the method, which is one important part of the optical character recognition (OCR) system.

In this paper, we present a novel 3D barcode preprocessing technique used in tracking the metallic honeycomb catalyst, which is a part of the automobile, on the manufacturing process of automatic detection. Because the parts need identifying before and after routinely high-temperature heating treatment stage, the information of the online system in each stage varies largely together with big disturbance. Thus we identify 3D barcode on parts with an industrial camera under an auxiliary illumination. Firstly, applying the method of image histogram equalization to enhance the contrast between bar and space, then transforming it by projection, next getting the bar and space information with self-adaptive threshold curve, finally detecting the correct information and gaining the parts' serial number decoding according to the national standard.

In this paper we present improved training algorithms to two newly developed classifiers, reduced set vector machines and Adaboost cascade classifier applied in face detection, which are all based on learning from data. Support vector machine (SVM) has been proved to be a powerful tool for solving practical pattern recognition problems based on learning from data. Due to large number of support vectors learnt from huge amount of training data the SVM becomes too computational intensive to many critical problems. Reduced set vector machine (RVM) is a faster approximation of SVM, but calculate a RVM is very difficult. Cascade classifier using Adaboost is a newly proposed method, which is much faster than the SVM and MLP methods and very competitive in performance to the existing systems , but the training is not easy due to feature number option. Our improved training algorithms make training become easier and more reliable and applicable in practice.

A feature based approach cooperating with Evolution Strategies for image correspondence estimation is proposed. As an optimization algorithm, Evolution Strategies is employed to search the optimal correspondence through out the source images. Instead of finding the correspondence of the entire image, the spatial relationships ofthe feature configuration in each image are discovered, where feature configuration is defined as the cluster of feature vectors on an image representing homogeneous feature distribution. Employing ES reduces the computational expense and broadens the applicable area since the comparison is restricted to the area inside the search structure of Evolution Strategies, which is defined as an ellipse. The results from various test images prove it to be an efficient and effective approach.

A counter-propagation network (CPN) based system of multi-sensor data fusion at feature level for target classification is proposed in this paper. This presentation mainly describes the use ofthe CPN in the data fusion system for target classification, as well as the algorithm used for training the CPN. As a demonstration of the advantages of the CPN, a popular back-propagation network (BPN) and a standard counter-propagation network (SCPN) are investigated at the same time. Finally, to illustrate the effectiveness ofthe CPN with the modified training algorithm for data fusion at feature level, we present the experiments for the application system based on FUR and TV camera. The experimental results for the system using the real-world database show that the CPN with the proposed algorithm provides the best overall performance. The classification accuracy, robustness and learning process are significantly improved.

We present a range of image processing tecimiques as potential pre-processing steps, which attempt to improve the performance of the eigenface method of face recognition. Verification tests are carried out by applying thresholds to gather false acceptance rate (FAR) and false rejection rate (FRR) results from a data set comprised of images that present typical difficulties when attempting recognition, such as strong variations in lighting direction and intensity, partially covered faces and changes in facial expression. Results are compared using the equal error rate (EER), which is the error rate when FAR is equal to FRR. We determine the most successful methods of image processing to be used with eigenface based face recognition, in application areas such as security, surveillance, data compression and archive searching.

In this paper, we propose an approach based on Kolmogorov Complexity (KC) measuie for determining script classes in mixed Chinese (complex characters)/English document images. This approach, which mainly consists of two steps: document image preprocessing and KC measure, can successfully separate Chinese text lines from English ones. Our approach is robust and reliable in handling document images of different appearances and densities, and various fonts, sizes and styles of characters used in documents. Experimental results on a set of 40 text line images (20 English text lines and 20 Complex Chinese text lines) from various document images show that 100% correct classification rate can be achieved.

In this paper, an automatic and real time method for detecting eye states is proposed. The method is based on the fact that the regions of iris and white regions of an eyeball can be detected when it is open. In this case, color images contain more information helpful for detecting irises than intensity images. The saturation of color is used to detect whether the eye is open or closed, then the edge map of the eye image is used to detect the irises. Using the color information to detect the locations of irises is more accurate than using only the edge image. The image sequences are nearly frontal-view color facial image sequences. In the paper, some eye sequences, which are detected by this method, are shown.

Fiber identification has been a very important task in many industries such as wool growing, textile processing, archaeology, histochernical engineering, and zoology. Over the years, animal fibers have been identified using physical and chemical approaches. Recently, objective identification of animal fibers has been developed based on the cuticular information of fibers. Effective and accurate extraction of representative features is essential to animal fiber identification and classification. In the current work, two different strategies are developed for this purpose. In the first method, explicit features are extracted using image processing. However, only implicit features are used in the second method with an unsupervised artificial neural network. It is found that the use of explicit features increases the accuracy of fiber identification but requires more effort on processing images and solid knowledge of what features are representative ones.

Texture plays an important role in image analysis. There are mainly two categories of texture analysis techniquesstatistical methods and structural methods, both of which have their own advantages. Based on the two categories, a novel approach is proposed, which describes linear texture using a kind of ARG (Attributed Relational Graphs), and is applied to image retrieval using probabilistic relaxation algorithm. Because the image 's characteristics are presented intuitively through structural attributes of graph, this method has excellent ability of texture description. Furthermore, straight-line segment is extracted as texture primitive. Three geometrical relation attributes between texture primitives, which are invariant to changes of scale, rotation and translation, are put forward. Therefore the applications of image retrieval basing this method possess excellent anti-noise ability. Practical experiments show that the performance of the new approach is satisfying.

This paper presents a novel algorithm for automatic localization of human eyes in grayscale still images with complex background based on geometrical facial features and image segmentation,. First of all, a determination criterion of eye location is established by the priori knowledge of geometrical facial features. Secondly, a range of threshold values that would separate eye blocks from others in a segmented facial image is estimated from the facial image histogram. Thirdly, with the progressive increase of the threshold by an appropriate step in that range, the size of the existing blocks in the segmented facial image will expand, some existing blocks will merge into one block, and some new blocks will emerge. Once two eye blocks appear from the segmented image, they will be detected by the determination criterion of eye location. Finally, the 2-D correlation coefficient is used as a symmetry similarity measure to check the factuality of the two detected eyes. In this way, the optimal threshold value can be automatically found based on the result of detection such that eyes can be accurately located. The experimental results demonstrate the high efficiency of the algorithm in runtime and correct localization rate.

In this paper, we present a natural gesture extracting method integrating arm edges detection, skin color model and cloth texture feature. We use skin extraction result to guide arm edge locating and use arm edge information to verify skin extraction result. Thus, some small skin regions in background are discarded. Our method can cope with some special cases such as two hands placed together, and also distinguish left and right hands. Further more, we add texture feature of cloth into arm position evaluation. Now the requirement for corsetry or marks is no longer needed.

Fisher discriminant methods (FDM) have been demonstrated their success in face recognition, detection, and tracking. Fisher discriminant method is based on the optimum of Fisher discriminant criterion. Recently Higher Order Statistics (HOS) has been applied to many pattern recognition problems. In this paper we investigate a generalization of FDM, Kernel Fisher discriminant methods (KFDM), for the feature extraction of face images, which is nonlinear analysis method. In conventional FDM, all the matrices including within -class scatter matrix, between-class scatter matrix and population scatter matrix are actually a second order correlation of patterns respectively, KFDM provides a replacement which takes into account ofhigher order correlation. Further more, KFDM computes the higher order statistics without the combinatorial explosion of time and memory complexity. We compare the recognition results using KFDM with conventional FDM on ORL face image database. Experimental results show that the proposed KFDM outperforms conventional FDM in face recognition.

Fingerprint enhancement is a key issue in fingerprint minutiae extraction due to various image qualities. Most of exist fingerprint enhancement algorithms need to estimate the ridge orient and ridge distance, which are used to design the enhancement filter. However ridge distance estimation is a rather dffIcult task due to the fingerprint image quality and singularity, therefbre, the ridge distance estimate usually fails in those case and leads to enhancement algorithm failed. In this paper, we propose an AM-FM basedfingerprint image enhancement algorithm, which uses a novel Dominate Component Analysis (DGA,) technique to estimate the dominate component, and uses band pass filter to enhance those component rather than directly estimate the ridge distance, ridge orientation. Experiment results show that our enhancement algorithm leads to signflcant image quality improvement as well as the system efficient.

In this paper the problem of training Support Vector Machines (SVMs) for video basedface recognition is presented. Faces as training samples are automatically extractedfrom input video sequence by multiple related template matching and normalized both in geometry via ffIne transformation based on corresponding facial feature points detected in the Sobel convolvedface regions and in gray level distribution via linear transformation to the same average and squared difference. Two different strategies for q-class face recognition problems with SVM are discussed both for ensemble face f eature set andfor PCA compressedfeature set. The performance ofa prototype system based on this technology over 100 clients is reported to demonstrate its greatpotentials in future.

In this paper, a direction sequence string matching based on-line signature verification system is proposed. A signature is coded as a direction sequence string. The modified edit distance is used for string matching. The test signature is compared with 5 reference signatures and distance is given by averaging the 5 distances. A verification result is given by comparing the distance measure with a pre-calculated threshold. The experiment shows a result of 4.7% equal error rate (EER).

This paper proposed a method for extracting Chinese characters in a gray scene image. This method makes use of the general features of. Chinese characters in scene image. First, the scene image is divided into partial regions. Some regions can be set as candidates of character regions with comparing spatial frequency and contrast. Then, the whole image is binarized using the dynamic thresholding method of gray-level image. Next the circumscribed squares of each candidate character regions are located according to the labeling results. Finally, the character regions are determined using the proximity of characters, co-linearity of characters and similarity of characters. The extraction rate is about 80.6% according to the experiments.

Colorectal cancer has been increasingly becoming one of the major killer diseases in the world. If this type of cancer can be detected in the early stage, the rate of survival is quite high. However, generally it is detected considerably late due to a lack of accurate screening program in practice for early detection. In this paper, a novel approach based on decision fusion is presented for providing an intelligent computer-assisted clinical diagnosis to physicians for the early detection of colorectal cancer. There are three steps in the screening procedure, that is, image acquisition from endoscope, image processing, and decision-making based on fusion technique. In the second step, many effective image-processing techniques can be applied and corresponding local decisions can be made for diagnosis. These decisions are fused in a fusion model to achieve the fmal decision. The fusion algorithm takes into account the accuracy of each technique and the performance of individual technique is utilized to adjust the weights employed m the algorithm. The proposed fusion approach is trained and tested by sets of endoscopic images. The results obtained are encouraging and suggest that the new fusion approach for the cancer detection is feasible.

A fundamental issue in 3D model-based vehicle tracking is pose recovery consisting of pose evaluation and refmement. In this paper, we compare two algorithms for pose evaluation. For the sake of self-containedness, the algorithms are first outlined. Experimental results under various conditions are then shown. We compare and analyze the results, and also discuss possible improvements ofthe two algorithms that may lead to better performance.

In this paper, a novel algorithm is presented for writer identification from handwritings. Principal Component Analysis is applied to the gray-scale handwriting images to find a set of individual words which best characterize a person's handwriting style and have maximal difference from other people style. During identification, we only need to utilize a set of individual characteristic words for comparison, instead of comparing the whole handwriting text to identify the writers. So not only is a very high average identification performance of 97.5% obtained, but also a very fast identification speed is achieved in our method. In the experiment, 400 pages ofhandwriting texts, containing almost 16000 Chinese words written by 40 different writers are used to validate the performance ofthe method.

In this paper we explore how to embed symbolic relational graphs with unweighted edges in eigenspaces. We adopt a graph-spectral approach. The leading eigenvectors of the graph adjacency matrix are used to define clusters of nodes. For each cluster, we compute vectors of spectral properties. We embed these vectors in a pattern-space using principal components analysis and multidimensional scaling techniques. We demonstrate both methods result in well-structured view spaces for graph-data extracted from 2D views of 3D objects.

The wavelet theory has become hot in the last few years for its important relative characters, such as, subband coding, multiresolution analysis and filter banks. In this paper, we propose a novel method of feature extraction for palmprint identification based on wavelet transform, which is very efficient to handle the textural characteristics of palmprint images at low resolution. The matching results show that the proposed feature extraction method is efficient in terms of matching accuracy and computational speed.

This paper discusses Moise cable transmitters' manual dexterity differences using pattern recognition. Firstly, using amplitude and frequency filtering of the frequency sampling voice signals results in the extraction of the characteristics of demodulated base band signal. Secondly, the calculation of the minimum space distance allows recognition of the manual techniques differences of Moise cable transmitters.

This paper presents a multichannel Gabor filter scheme for on line textile defect detection. A textile image is processed by using the Gabor-filters with multiscale and multioriention and hereby form multi-images. Then, the images are reconstructed into one image for detecting defects. The experimental result shows that the algorithm is robust and computationally efficient. The overall detection rate is around 95%.

The traditional methods of lens distortion correction need man inference and don't be directly computed. The methods need ether calibrated reference objects or the correspondence information that men input. The paper provides the way that can directly compute lens-radial distortion parameter. By means of alignment multi frames with rotating motion model, the way can correct the radial lens-distortion parameter. The paper gives a model and trial results. It deduces formulates. The trial results show the good correction for lens-distortion images.

A method for summarizing the information of a video clip into a single image is proposed. Two kinds of information in video clips can be distinguished, one is related to background contents and other is related to foreground object motion. For condensing foreground object motion information, a new technique based on edge overlap is described. The edges of moving objects in a set of consecutive frames are detected and then suitably overlapped on the composite background to show the movement of objects along time axis. The background variation caused by camera motion is captured and different scenes are connected using video mosaic technique. By combined use of video mosaic and edge overlap techniques, a VM&EO frame is generated, which sums up both the background contents and object motion information. Thus, such a frame can be used to represent the video clip in a compact and meaningful way. This will greatly save people's time for viewing the whole video clip to capture the necessary motion information in the video browsing and retrieval systems as well as other applications.

This paper discusses a method of the conjunction of the neighboring orbit satellite image data. It uses the plane method and surface method to search the conjunction area, and gives a mathematical model to judge the conjunction pattern. According to the step, which can connect the neighboring orbit satellite data, it gives a whole process and three methods ofconjunction. (Left Image Method, Right Image Method, Balance Image Method)

This paper describes new algorithms for detecting and reconstructing the road with a single view for highway application. We concern about not only the whole road but also the white lines on the road. In real navigation, the white lines are more important than the whole road. By processing the gray images in real time, which captured by a CCD, we get the road's location approximately and extract the white lines exactly. In order to direct the navigation accurately and detect the obstacle fast, we use the conic model to reconstruct the white lines. In this paper, we present how to extract the road feature points and how to determine the conic model parameters. We experimented our system with the algorithms on highway at 120km/h in Sichuan province and Chongqing city. The result shows our system can work perfectly.

An approach which uses multiple sources of visual information (or visual cues) to identify and segment the ground plane in indoor mobile robot visual navigation applications is presented. Information from color, contours and corners and their motion are applied, in conjunction with planar homography relations, to identify the navigable area of the ground, which may be textured or non-textured. We have developed new algorithms for both the computation of the homography, in which a highly stable two point method for pure translation is proposed, and the region growing. Also, a new method for applying the homography to measure the height of a visual feature to the ground using an uncalibrated camera is also developed. Regions are segmented by color and also by their sizes and geometric relation and these region boundarys are extracted as contours. By controlled manoeuvres of a mobile robot, the methods of coplanar feature grouping developed in this paper are not only applicable to corner correspondences but also to contours. This leads to robust, accurate segmentation of the ground plane from the other image regions. Results are presented which show the validity of the approach.

In this paper, we introduce the linear constrains of circular points on camera's intrinsic parameters firstly and propose the uniqueness condition of the solutions, then we elaborate a novel linear approach for computing the images of circular points as well as camera calibration from images of two unparallel coplanar rectangles in space. The main advantage of our technique lies in that neither the metric measurement of the rectangles nor the correspondences between images are required. Extensive simulations and experiments with real images, as well as the comparative study with Zhang's method, validate our algorithms and demonstrate that the proposed technique is of high precision and strong robustness.

Affine reconstruction is the most difficult and crucial step in 3D reconstruction from image space to the Euclidean space. Recent research indicates that it is impossible to realize the affine reconstruction from a perspective image pairs captured by a translating camera with varying intrinsic parameters if no geometric information of the scene is available. Therefore, it is indispensable to provide some additional information for affine reconstruction from a pair of images. In this paper, we propose that if one plane and a pair of parallel lines are presented in the scene, the affine reconstruction can be done linearly from two images taken by a translating camera. We also point out that if a pair of parallel planes is presented in the scene, the affine reconstruction can also be done linearly from an image pair taken by a translating camera, and if a pair of parallel planes and a pair of parallel lines are presented in the scene, the affine reconstruction can be done linearly from two general views. Extensive simulations and experiments with real images validate our algorithm. The result ofthis paper seems ofboth academic and practical significance in 3D computer vision.

Image-based rendering technique is a very interesting research area in Virtual Reality. This paper presents a novel approach to create spherical panoramas. With these panoramas, we make use of the idea of hyperlink organizing viewpoints and the virtual objects to construct image-based spherical virtual spaces. In such spaces, users can look in all directions in one viewpoint, wander widely by jumping from one viewpoint to another, and manipulate virtual objects freely.

Extensive researches have been executed to extract the camera motion in videos by utilizing temporal slices, analyzing optical flow distribution, using transformation model, etc. However, these strategies fail to detect the camera rotation; furthermore, extracted optical flow or motion vectors may contain considerable noise or error, which significantly reduces the efficiency of these strategies. In this paper, the mutual relationship between motion vectors is utilized for qualitative camera motion classification. We first define four types ofmutual relationships (parallel, approach, diverging and rotation) between any two motion vectors, then, a 14-bins feature vector is constructed to characterize the statistical motion information for each P-frame. Based on different distribution modes ofthe motion feature vector, the qualitative camera motion classification is executed. In addition to detecting most common camera motions (pan, tilt, zoom, still), our method can also detect camera rotation. Experimental results demonstrate successful classification over different types of video collections.

In this paper, we propose a novel scheme for matching planar images with three-dimensional geometrical shapes. Given a two-dimensional unknown image and a 3D model, Simple Genetic Algorithm is applied to search for the existence of a legitimate projective transform that can spatially associate the two entities. A positive outcome reflects that the image is possibly one of the many views ofthe model. To enhance the search efficiency, an alternative representation of the projective transform has been adopted so that the search space can be reduced. The method had been successfully applied to identify unknown views amongst a moderate collection of 3D models.

Generating realistic 3D human face models and facial animations has been a persistent challenge in computer graphics. An automatic system that constructs continuous 3D facial animation models equipped with left-and-right cameras is developed. The facial features are automatically located, registered and partitioned with very dense triangular mesh. Each planar mesh point is associated with a depth value and a motion vector. A 3D geometry can be constructed by combining the planar mesh points with corresponding depth values, while the association of mesh points with displacement map produces animated facial expression. A 2D facial image of any character can follow the same feature extraction, registration and partitioning paradigms to establish the mesh point correspondence with the generic sequence stored. By inheriting the depth values and displacement map, an animated facial sequence of different person with the same motion characteristics is generated.

Image stabilization can be used in a variety of tasks in which dynamic image analysis is required, such as video steering, robot video monitoring, object tracking and wearable system. In this paper, we present a motion model to estimate the motion parameters for image stabilization by using feature points extracted from consecutive frames. Motion compensation is achieved by warping the current frame to a reference frame whose selection depends on different compensation methods. The experimental result shows that the stabilization rate is up to 25 frames per second with tracking window size 1 12x 128 pixels.

This paper presents an automatic soccer annotation system. The system consists of 3 modules: 3D reconstruction module, behavior analysis module and annotation module. The first module reconstructs the 3D model of the soccer scene, including positions, velocities and acceleration of each player, and the trajectory of ball. According to these parameters and domain knowledge, behavior analysis module will recognize each player's action with a finite state machine. Then, annotation module will convert key player's behavior into annotation words. The system proves to be robust when errors of 3D reconstruction results are small.

Traditional methods of stereovision computing need two or more images of the target obtained from precalibrated binocular or multiple cameras. But in real applications, the internal and external parameters of camera may change from time to time. So recently, uncalibration methods of stereovision computing gain more and more attention. The subject ofthis paper is 3D vision computation with uricalibrated cameras. Consider that in many robotic and industrial applications, the environment is structural and the knowledge about operation targets is known. This prompt us a possible solution of avoiding camera calibration. The proposed method in this paper can determine the position/pose of an object with some known geometry and measurements only from a single image even when the focus length is unknown. In this paper, the computation formulas for both binocular and monocular systems are derived with illustrations. Then the experimental examples are given, and analysis on the result data and errors is also presented. Experiments show that the computation is very simple and the result can be very accurate. Another advantage is that it can be used in monocular systems. The proposed method can be widely used in practical applications, provided the required features can be precisely obtained.

A factorization method is proposed for recovering camera motion and object shapes from line correspondences observed in multiple images with perspective projection. For any factorization-based approaches for perspective image, just like from point correspondences, scaling parameters called projecting depths must be estimated in order to obtain a measurement matrix that could be decomposed into motion and shape. One possible approach, proposed by Sturm and Triggs[Sturm-96], is to compute projective depths from linear relations of multiple images, for example, fundamental matrix for two images, and the trifocal tensor for three images. However, the estimation process of the fundamental matrices or trifocal tensor might be unstable if the measurement noise is large or the camera and the object points are nearly in critical configurations. In this paper, we first develop the line imaging geometry from the point imaging geometry, and then form a measurement matrix j ust like for the points, and propose an algorithm by which the projective depths are iteratively estimated so that the measurement matrix is made to be as close as possible to rank 6. At last, we factorize the measurement matrix into two matrices, and exploit the relation of Plucker coordinates of the lines to achieve the camera motion and shape in 3D projective space. This estimation process requires no the linear relations between the images and is therefore robust against measurement noises. The validity of the proposed method is confirmed by experiments with synthetic and real images.

The paper describes a new method of self-calibration of a moving vision platform (a rig) when the cameras' centers relative to each other during the motion have been determined. Given the projective structure from the corresponding points across the set of images acquired by the cameras, the varying internal and external parameters can be computed using only the positions of the cameras ' centers. The method in this paper shows that the maximal generality is reached when the rig consists of 5 cameras whose centers of projection are given. Furthermore, if the only goal is to construct the Euclidean structure from a projective structure, the internal and external parameters need not be computed, and the process is linear and hence it is more convenient and stable.

We present a system for tracking facial expression and head pose with one camera and no special markers, and generating a face animation to imitate these expressions simultaneously. The tracking is based on Gabor wavelet coefficients (jets) of the facial feature points, a saccadic searching strategy, and a model-based adjustment. Using Principal Components Analysis (PCA), we established an expression model and use it to drive the imitation animation. The system runs in real-time and the tracking shows robustness against illumination/background variation, accumulative error and partial occlusion. The imitation animation captures both the expression and the head pose without 3D information.

We iniroduced a web-based assembly/disassembly system in this article, which is designed for e-commerce and teleeducation purpose. The system is characterized by automatic disassembly pathlsequence generation, multi-level deployment/rendering, partially ordered reliable delivery and a friendly web-based user interface.

This article presents a new type of 3D color digital imaging system. First, we briefly introduce the current 3D imaging technologies: passive and active sensing. Second, a new type of 3D imaging system is introduced. Using this system, the shape of an object can be digitized and the texture of this object surface can also be gained simultaneously. The emphasis of the paper is put on the description of hardware design and software framework design of the system. The paper also presents the 3D digital image made on a human face as an experimental result. Then, the potential industrial applications such as reverse engineering, digitization of museum artifacts, inspection, biomedical imaging, home shopping, film-making and virtual reality, etc., are described. Finally, we make conclusion and future development on 3D imaging.

The problem of surface reconstruction from unorganized points has been, and continues to be, an important topic of research. Surface reconstruction can be widely used in reverse engineering and visualization of scientific data, etc. In this paper, a new algorithm for surface reconstruction from unorganized points in R is proposed. The algorithm is based on the theorem that the tangent plane of any point on a manifold surface local linear approximates the surface, which means that any point in one scattered point's neighborhood can be found one and only one projection in its tangent plane. In order to reconstruct a surface interpolating the scattered points, we first project the neighbor points of one sample point p to its tangent plane and find its 2D starlike Delaunay neighbors. After that, we define the points whose projection is 2D Delaunay neighbors of p as its 3D Delaunay neighbors. At last, the triangular mesh can be obtained based on the principle that three points consisted one triangular plane patch ifthey are 3D Delaunay neighbors each other. Experimental results show that this algorithm is effective, robust and the output mesh accords with Delaunay character.

Medical image elastic registration is an important subject in medical image processing. Previous work has focused how to select the corresponding landmarks manually and then use adequate interpolating for gaining the elastic transformation. However, the landmarks extraction is always prone to error, which could influence on the registration results. And localizing the landmarks manually is also difficult and time-consuming. We used Multiquadric method with smooth character, thereby , utilized a semi-automatic method to extract the landmarks .Combining these two steps, we proposed an accurate, fast and more robust registration method ,and obtained the satisfactory results.

In reverse engineering, a reconstructed surface model generally is composed of several surfaces with d continuity. With respect to the low quality surfaces, the geometric continuity along the boundary may be destroyed after fairing operation is applied. So as to rebuild a global d continuous model, a series of complex operations such as bridging, blending and filling are always inevitable. In this paper, a new fairing method, which is a B-spline surface fitting optimization subject to complicated boundary constraints, is presented. With this method, the users can avoid the complex surface editing work but focus on the surface quality. As a result, the final surface model will be with a higher quality while preserving the original geometric continuity as before. Some experimental results show that the above method is feasible and satisfying.

Taking use of VR, GIS and WWW technique, this paper discusses how to develop a software system, combining with theories of urban planning. This system could provide a platform for urban pkmning, on which the effect of design, environment and layout can be examined before construction. At the same time, the system could provide designing thta and special analysis. So the superintending departments and experts could evaluate the design in the system, and the infoimation of evaluation could be feed back to the system to redesign, till the result is satisfying.

An algorithm of LOD automatically generation through sparsely resampling technique is presented in this paper. Based on DEM uniform grid model, "block" is taken as the space unit of terrain simplification and a multi-LOD data structure is built to store every level of models. The complexity of calculation is greatly reduced, the rendering of the terrain model is simplified effectively, and the efficiency of the terrain generation is improved. Additionally, the effects of viewpoint and view direction are considered, which makes the 3D terrain more realistic. On the basis of the theory of this algorithm, a system for 3D complex terrain generation is realized through Visual C++6.O and OpenGL programming. The dynamic, real-time and visual characters of this system are satisfying, which verifies the validity ofthe algorithm further.

Combining the concept of weights in rational curves with singular blending technique, we have generalized Bézier curve to a generalized Bézier curve denoted as a -Bézier curve. Its shape-control capability is much better than that of Bézier curve; thus, ? Bézier curve is more useful in free curve and surface designing. Bézier curve can be converted into an a Bézier curve by adding blending parameter to control vertices of Bézier curve. The properties of a Bézier curve are studied in details, and the effects of the blending parameters are investigated. By varying the blending parameters the curve can be reshaped, so it is brightly useful in the applications of CAD/CAM.

Several approaches have been proposed to visualize 3D volume data on multiprocessor systems in recent years. Most of these algorithms are tailored for special architectures and hard to immigrate to new architectures, such as Tsinghua High-Performance Cluster System (THPCS), a distributed-memory multiprocessor system. In this paper, a parallel volume rendering algorithm is presented for THPCS. To reduce the communication cost and the complexity of blending local intermediate images, a static subtask distribution strategy is proposed. Combined with the prediction for the subtask load and node performance, our strategy yields a good load balance. An asynchronous binary compositing method is adopted to provide flexibility for our algorithm. The experimental results demonstrate the efficiency and practicality of our algorithm. As a result, THPCS using our algorithm achieves performance ofrendering 512 x 512 color image for 128x 128x 197 volume within 1 second.

A piecewise quadratic trigonometric rational spline curve is presented and analyzed in this paper. Each curve segment is constructed by using three consecutive control points. The spline curves can be C3 continuous for uniform knot vector and C2 continuous for non-uniform knot vector. A local shape parameter is used in the spline curves. As the increase of the value of shape parameter, the curves approach to the edge of the control polygon.

In interactive environments including computer games and virtual reality applications, we have increased need for interactive control of articulated body motions. Recently, physically based methods including constrained dynamics techniques are introduced to this area, in order to produce more realistic animation sequences. However, they are hard to achieve real-time control of articulated bodies, due to their heavy computations. In this paper, we present a procedural method for interactive animation of articulated bodies. In our method, each object of the constrained body is first moved according to their physical properties and external forces, without considering any constraints. Then, the locations of objects are adjusted to satisfy given constraints. Through adapting this two-stage approach, we have avoided the solving of large linear systems of equations, to finally achieve the interactive animation of articulated bodies. We also present a few example sequences of animations, which are interactively generated on PC platforms. This method can be easily applied to character animations in virtual environments.

This paper presents a new method of non-rigid registration based on elastic models incorporating shape information. Given a source and a target image, and the shape of interest in these two images, the idea is to incorporate the geometrybased shape knowledge into extemal forces, which will drive the deformation from the source image to the target image. The transformations are constrained to a physical model of elasticity to maintain smoothness and continuity. Some experiments were performed on both synthetic and medical images of brain.

Human walking is frequently simulated in Computer Animation, for it is the haic and widely used locomotion of human being. Many algorithms have been designed, which can be divided into two kinds: Kinematics based and Dynamics based. This paper concentrate on the simulation of curved path human walking in Virtual Environment. We use a new method, parabolic fitting, to optimize path generated by path planning. It is very convenient to get face direction and path length information from this method. A blend factor is provided to control the approximation level to original path of the optimized path. Inverse kinematics is then used to generate walking animation. The foot of stance leg is constrained to fix on the ground and the foot of the swing leg is controlled by a swing curve. Only high level parameters such as speed, stroke are needed to generate walking animation.

The design of Functional Graded Material (FGM) and the visualization relative technology are described briefly in this paper, such as the organization ofthe 3D regular or irregular data ,isoline and the isosurface technique. It is proved that the visualization is used in the design of Functional Graded Material (FGM),which makes the material design is more intuitive and more reasonable.

The triangular irregular network (TIN) method is a rational choice for building digital elevation model (DEM) to keep model more precise. The common problem in forming standard delaunay triangle mesh is time-consuming, especially with the large data of sampling nodes. Accordingly, we introduce a practical algorithm to reduce sampling quantity of terrain in a single delaunay triangle mesh and to increase the efficiency of building the mesh greatly. We tested the algorithm with several real sampling data with different distributions. We also present the comparison result of the proposed method with the traditional.