The gradient operator is frequently used for detecting edges in digital images. Although it does enhance edges well, it also enhances noise well, making edge detection difficult, especially for computer vision systems. This paper introduces a simple improvement of the gradient operator called gradient propagation. The standard method of detecting edges with a gradient operator is to calculate the gradient at each pixel and then threshold the magnitude of that gradient. Gradient propagation does exactly the same thing except that before thresholding, it adds copies of the gradient vector at each pixel to an output gradient image along a line perpendicular to the gradient vector. The gradient sums turn out to be large along edges and small noise. After thresholding, strong image edges processed by gradient propagation come out as good as those from the gradient operator but weak edges are detected better and with substantially less noise. The paper compares the results of the Sobel operator and equivalent edges. Both techniques produce little noise when detecting strong edges, but for weak edges, gradient propagation shows substantially less noise.

An innovative method that enhances detail in digital images by smoothing image pixels while introducing minimal distortion is described and tested. In particular, a 14 by 14 pixel region of a diital image is smoothed using a constrained Gaussian radial basis function method. This method centers on each pixel a Gaussian distribution of amplitude such that the sum of all distributions correctly reproduces the gray level of each pixel. To assess the method, the distortion of the smoothed image is measured by the deviation of its power spectrum, from that of the unsmoothed image, determined as a function of the Gaussian distribution width, and comparisons are made with bilinear interpolation, a conventional convolution smoothing technique. The new method is capable of removing more 'pixel noise' while introducing less image distortion, thus permitting the detection and examination of otherwise hidden detail in digital images. Examples include the detection and assessment of enemy weapons in military images and cancerous tumor medical images.

It is unfortunate but true that incomplete imaging system models are commonly used for simulation-based image restoration studies. As an illustration of this, a common-but- incomplete continuous-input/continuous-output (c/c) system model is used in this paper to produce some simulated restorations. In this way, it is demonstrated that when image formation blur is the only significant source of image degradation, then conventional c/c model-based restoration can successfully sharpen blurred images, even when the blurring is excessive. If however, at least a small amount of additive random noise is also a source of image degradation (as is always the case in practice), then conventional c/c model-based restoration will produce satisfactory results, if at all, only if the restoration filter is consistent with a more comprehensive system model that accounts for the presence of this noise. Moreover, if sampling and reconstruction are part of the imaging process then conventional c/c model-based restoration can fail to produce satisfactory results; this is true even if there is no additive random noise. In this way, it is demonstrated that a c/c model is not a correct sampled imaging system model; instead a more comprehensive continuous-input/discrete- processing/continuous-output model should be used as the basis for simulation-based restoration studies.

Warping similarity transformations provide a powerful vehicle for generating new classes of joint distributions based on concepts different from time, frequency, and scale. These new signal representations focus on the critical characteristics of large classes of signals, and hence, prove useful for representing and processing signals that are not well matched by current techniques. Interestingly, all distributions that have been used to illustrate more complicated generalized distribution design techniques can be generated using the warping method.

As introduced by Matheron, granulometries depend on a single sizing parameter for each structuring element forming the filter. The size distributions resulting from these granulometries have been used successfully to classify texture by using as features the moments of the normalized size distribution. The present paper extends the concept of granulometry in such a way that each structuring element has its own sizing parameter and the resulting size distribution is multivariate. Classification is accomplished by taking either the Walsh or wavelet transform of the multivariate size distribution, obtaining a reduced feature set by applying the Karhunen-Loeve transform to decorrelate the Walsh or wavelet features, and classifying the textures via a Gaussian maximum-likelihood classifier.

The projection of structured light is a technique frequently used in computer vision to determine surface structure of scene objects. In this paper, higher level features are extracted from the images and used for a direct estimation of second-order object surface models. The algorithm is based upon a predictor-corrector approach which utilizes an initial estimate for the surface parameters, followed by iterative parameter refinement. A predicted passive image is generated using the current surface parameter estimates and significant features are extracted and compared with those in the true passive image. The extimated surface parameters are corrected based upon feature disparities. The algorithm is well-suited for a particular vision task involving recognition of cylindrical drums. In computer simulations and laboratory experiments, the algorithm was found to converge quickly and to yield accurate results.

Granulometric spectral decomposition results from partitioning an image according to the manner in which a granulometry continuously diminishes the image. A granulometric bandpass filter operates by passing some components and not passing others. An optimal granulometric bandpass filter is one that passes components in a way to minimize the expected area of the symmetric difference between the filtered and ideal images. The present paper considers bandpass optimization for reconstructive granulometries. For these, each connected grain in the input image is either fully passes or eliminated. Hence, such filters are well-suited to elimination of clutter. The observed image is modeled as a disjoint union of signal and clutter grains and the filter is designed to best eliminated clutter while maintaining the signal. The method is very general: grains are considered to be realizations of random sets; there are no shape constraints (such as convexity) on signal and noise grains; there are no similarity constraints between granulometric and image generators, and the method applies to overlapping grains by considering the filtering to take place on the image model resulting from segmentation preprocessing.

Most image processing and feature extraction algorithms consist of a composite sequence of operations to achieve a specific task. Overall algorithm capability depends upon the individual performance of each of these operations. This performance, in turn, is usually controlled by a set of a priori known (or estimated) algorithm parameters. The overall design of an image processing algorithm involves both the selections of the sub-algorithm sequence and the required operating parameters, and is done using the best available knowledge of the problem and the experience of the algorithm designer. This paper presents a dynamic and adaptive image processing algorithm development structure. The implementation of the dynamic algorithm structure requires solving of a classification problem at decision nodes in an algorithm graph, A. The number of required classifiers equals the number of decision nodes. There are several learning techniques that could be used to implement any of these classifiers. Each of these techniques, in turn, requires a training set. This training set could be generated using a modified form of the dynamic algorithm. In this modified form, a human operator interface replaces all of the decision nodes. An optimization procedure (Nelder-Mead) is employed to assist the operator in finding the best parameter values. Examples of the approach using real-world imagery are shown.

The problem of separation of binary stellar systems consists in reconstruction of flux and coordinates of each component from their image. To solve this problem the special algorithm is constructed, which is based on connection of stellar system geometry with the image intensity moments up to the third order inclusively. It is important to note that the form of the point spread function (PSF) is considered to be arbitrary. The only restriction to the PSF is its central symmetry. During processing it is convenient to find the canonical coordinate system, i.e., the coordinate system, connected with the intensity distribution. The distinction between the moments of the second order in the canonical coordinate system determines extraction of the stellar system in the image and the x-moment of the third order determines the asymmetry of the system. To find the characteristics of the stellar system, the set of simultaneous equations was derived. It expresses the moments of the intensity distribution through the geometry of the stellar system and the PSF. This set of equations is uniquely resolvable with respect to the parameters of interest, i.e. the flux and the coordinates of each component. It is worth noting that the final results does not depend on the PSF, provided it is centrally symmetric.

The discrete wavelet transform (DWT) has recently been proved a powerful tool for image compression because of its multiresolution decomposition aspect. However, this algorithm takes high computation time, so it is necessary to use a parallel computer to compress 6000 X 6000 satellite images in real time. This paper deals with the parallelization of the image compression algorithm based on DWT. We present implementation results on several parallel computers (CM5, CM200, OPENVISION, SYMPHONIE).

A visual communication channel can be characterized by the efficiency with which it conveys information, and the quality of the images restored from the transmitted data. Efficient data representation requires the use of constraints of the visual communication channel. Our information theoretic analysis combines the design of the wavelet compression algorithm with the design of the visual communication channel. Shannon's communication theory, Wiener's restoration filter, and the critical design factors of image gathering and display are combined to provide metrics for measuring the efficiency of data transmission, and for quantitatively assessing the visual quality of the restored image. These metrics are: a) the mutual information (Eta) between the radiance the radiance field and the restored image, and b) the efficiency of the channel which can be roughly measured by as the ratio (Eta) /H, where H is the average number of bits being used to transmit the data. Huck, et al. (Journal of Visual Communication and Image Representation, Vol. 4, No. 2, 1993) have shown that channels desinged to maximize (Eta) , also maximize. Our assessment provides a framework for designing channels which provide the highest possible visual quality for a given amount of data under the critical design limitations of the image gathering and display devices. Results show that a trade-off exists between the maximum realizable information of the channel and its efficiency: an increase in one leads to a decrease in the other. The final selection of which of these quantities to maximize is, of course, application dependent.

In this paper, we will first briefly review fractal equations, then we will present an algorithm for solving encoding equations of class 2 fractals. The idea is to represent images in terms of blocks in a similar way as JPEG. the DCT in JPEG is replaced by fractal transformation. Image compression ratios for various selections of block size and parameter space will be calculated. Thirdly, we will study the simplest IFS, a 1D IFS with two mappings. We will introduce several new concepts: the base language of fractals and the canonical order of fractals. The cardinality of the base language is derived. How base language is used to infer an IFS is discussed.

In this paper, we propose two new fast algorithms for motion vector (MV) estimation using spatial correlation of MVs in adjacent blocks. We select a set of MV candidates based on the MV knowledge of its neighboring blocks, and then perform further search to refine the MV result. The first algorithm is performed on each block consecutively. The second algorithm is a modified version of the first one by using a block subsampling technique to reduce computational cost. We show with experimental results that, compared with full search block matching algorithm (FBMA), the proposed algorithms have a speed-up factor ranging from 50 to 100 with only 2-15% MSE increase when applied to typical test image sequences.

A vast number of applications including defense, medical, manufacturing, law enforcement, digital library, education, space exploration, weather forecasting, and entertainment require efficient management of huge collections of nonalphanumeric data. The most common and important nonalphanumeric data in most of these applications is image data. Owing to the availability of a variety of visual sensors, several large collections of images and related anciliary data exists and are rapidly growing. Examples of such collections include LANDSAT, weather, medical, and DoD target signature images. Unfortunately, in most cases only a fraction of the collected data is ever utilized to its full potential. The primary reason for this under-utilization is the lack of pictorial data management techniques/systems. Conventional data management systems are not designed to handle pictorial data in an integrated fashion, i.e., images and alphanumeric data are not treated equally. In such systems, an image is stored as a tag field in the description of some entity. Images are not entities and they cannot be key fields. Furthermore, content-based retrieval of images and related data is not possible. Therefore, new data management technologies need to be developed for an integrated management of textual and imagery data. This requires a clear understanding of the requirements and desireable characteristics of a pictorial data management system. In almost all image information management (or integrated image database) applications, image information modeling, content-based image information retrieval, and memeory management are the most important issues to be resolved. In this paper, the requirements of an integrated image information management system and the challenges posed by image data from the data modeling, the content-based retrieval, and the memory management viewpoints are discussed.

A new approach of using the VQ codewords as the remote sensed image features for content- based retrieval is proposed in this research. Different distortion measures are tried in the VQ stage to enhance the performance of the codewords as 'content descriptors' including classification accuracy. A system based approach has been taken to ensure that the features satisfy the different criteria imposed by a whole system. We implemented two main types of queries--query by class and query by value. The performance with respect to the former query was satisfactory while that for the latter query was excellent.

An ideal image compression technique for image data archival, retrieval and distribution would be one with the asymmetrical computational requirements of vector quantization (VQ), but without the complications arising from VQ codebooks. Codebook generation and maintenance are stumbling blocks which have limited the use of VQ as a practical image compression algorithm. Model-based VQ (MVQ), a variant of VQ described here, has the computational properties of VQ but does not require explicit codebooks. The codebooks are internally generated using mean removed error and human visual system (HVS) models. The error model assumed is the Laplacian distribution with mean, (lambda) , computed from a sample of the input image. A Laplacian distribution with mean, (lambda) , is generated with a uniform random number generator. These random numbers are grouped into vectors. These vectors are further conditioned to make them perceptually meaningful by filtering the DCT coefficients from each vector. The DCT coefficients are filtered by multiplying by a weight matrix that is found to be optimal for human perception. The inverse DCT is performed to produced the conditioned vectors for the codebook. The only image dependent parameter used in the generation of codebook is the mean, (lambda) , that is included in the coded file to repeat the codebook generation process for decoding.

The limited channel capacity associated with telecommunication and data network necessitates applications of image data compression for teleimaging systems. The UNIDAC technology is based on the optical consolidation of lossy and lossless compressions controlled interactively by a remote image analyst. The UNIDAC technology incorporates the positive features of separate lossy (a high data compression ratio) and lossless (errorless image quality) image data compressions without their associated weaknesses. The high value of that data compression ratio achieved by the UNIDAC technology is based on the elimination of a positional statistical redundancy additionally to a spatial statistical redundancy, and on the sequential nature of the visual analysis. A positional statistical redundancy reflects the fact that the data essential for image analysis is enclosed not in the whole image, but rather in parts of it--i.e. the window (area) of interest. In the professional knowledge to locate the areas of interest within the lossy compressed image. Selected positional information for the window of interest is transmitted back to the image source. The lossless compressed/decompressed residual image data is then used to update the image in the window of interest to its original lossless, errorless image quality. The potential capabilities of the UNIDAC technology are illustrated by its application for such teleimaging systems as teleradiology, telepathology, telesurveillance, and telereconnaissance.

The paper focuses on the problem of the mulitspectral image segmentation. A multispectral image is composed by several monospectral images (black and white for example), constituted by different wavelength bands. Consequently, the complementarity and/or redundancy of data, through data fusion, makes reliable and robust military or individual systems. The parallelism of the algorithm is inherent: different monospectral images may be sometimes processed in parallel without information exchanges or synchronization, or different monospectral images may be processed in parallel, but with explicit information exchanges and synchronization. This method is a new approach of image data fusion. It does not enter the taxinomy of image data fusion methods, because its level of fusion is between two classical levels. The image data fusion is performed while segmenting together the different images. The image data fusion is treated as an extension of segmentation methods. The classical or monospectral image features edges and regions that have been extended to the multispectral framework. Their attributes gather the information coming from all spectral images.

Vehicle-tracking is defined as an autonomous vehicle following a lead vehicle based on the observed range and heading angle of the lead vehicle. This paper presents a vision-based vehicle tracking algorithm for obtaining the range and heading angle information. The developed algorithm utilizes the minimization of an energy function reflecting the optical flow of a specified shape on the lead vehicle and the gradient of its edges. The embedded optical flow method is capable of tolerating intensity variations or environmental noise generated by light and background sources. This noise tolerance is achieved by employing a response function which adapts to the distribution of a similarity measure within a moving search window. Several examples are shown to illustrate the performance of this tracking algorithm in realistic outdoor scenes.

This paper concerns the determination of orientation of road vehicles from monocular intensity images. A novel algorithm is presented which exploits known physical and geometric knowledge about traffic scenes to allow fast and model-independent determination of vehicle orientations. The algorithm eliminates the need for symbolic image feature extraction and image-to-model matching, and the computational cost is substantially reduced. In fact, since the algorithm only requires local gradient data, object orientation can be determined from the input video data on-the-fly, and the overall algorithm can easily be implemented in real-time. The algorithm is tested with both indoor and outdoor data. Successful results are obtained for a variety of vehicles in routine traffic scenes.

In this paper, the possibility of driving egomotion in order to dock with a nonmoving target using information on the first order approximation of the optic flow in the center of the image is presented; specifically it will be shown how a camera mounted on the end-effector of a six degrees of freedom robot arm can be driven toward a planar surface whose position and orientation in space are unknown, avoiding collisions with the target and orienting the camera with its viewing direction perpendicular to the surface. It will be described how it is possible to achieve this task combining the effect of three simple independent processes: braking, fixation, and sideway translation.

Robot systems that rely on vision as their main sensory capability need to be able to cope with changes in the visual environment and to manage a wide field of view. Moreover, in order not to loose real-time response capabilities, selective visual sensing is indeed highly desireable. The 'built-in' selection in space and time provided by space variant sensors acts as a filter on the visual field having considerable implications for robotic applications. This paper focuses the attention on log-polar vision in the context of active control of visual sensors. The geometric distribution of sensing elements in the log-polar mapping provides visual task simplification and computational advantages. Correlation measurement techniques in the log- polar framework are formalized and two different uses are proposed. By performing global measurements on convergent log-polar images, a binocular mount can drive its cameras towards correct vergence configuration. It is also shown how image shifts can be detected by using 1D correlation measurements in the log-polar domain, and a possible use of this technique aimed at stabilizing gaze or tracking moving objects is presented. Both the reduced algorithm complexity, due to space variant topology, and the computational advantages, due to the limited number of pixels, make log-polar mapping a good candidate image geometry to obtain real-time responses in the context of reactive vision systems.

Two systems for velocity-based visual target tracking are presented. The first two computational layers of both implementations are composed of VLSI photoreceptors (logarithmic compression) and edge detection (difference-of-Gaussians) arrays that mimic the outer-plexiform layer of mammalian retinas. The subsequent processing layers for measuring the target velocity and to realize smooth pursuit tracking are implemented in software and at the focal plane in the two versions, respectively. One implentation uses a hybrid of a PC and a silicon retina (39 X 38 pixels) operating at 333 frames/second. The software implementation of a real-time optical flow measurement algorithm is used to determine the target velocity, and a closed-loop control system zeroes the relative velocity of the target and retina. The second implementation is a single VLSI chip, which contains a linear array of photoreceptors, edge detectors and motion detectors at the focal plane. The closed-loop control system is also included on chip. This chip realizes all the computational properties of the hybrid system. The effects of background motion, target occlusion, and disappearance are studied as a function of retinal size and spatial distribution of the measured motion vectors (i.e. foveal/peripheral and diverging/converging measurement schemes). The hybrid system, which tested successfully, tracks targets moving as fast as 3 m/s at 1.3 meters from the camera and it can compensate for external arbitrary movements in its mounting platform. The single chip version, whose circuits tested successfully, can handle targets moving at 10 m/s.

Log-polar pixel tessellation in the image plane improves binocular stereo perception compared to the familiar uniform Cartesian tessellation. This paper describes the advantages by analyzing the 3D intersections of projections of dual retinas. The 3D environment is divided into volume cells (voxels) which are the intersections of pixel projection cones. There are some interesting and useful differences between log-polar and Cartesian induced voxel distribution. Maximum stereo resolution for Cartesian voxels is inconveniently located at the outskirts of the field of view at the near point of intersection of the two fields of view, rapidly degrading therefrom. Log-polar stereo resolution is highest at the point of intersection of the optical axes. Active vision can steer this focus of attention like a spotlight to any point of interest in the 3D environment. Within this focus of attention, stereo resolution is nearly uniform. Applications include active robot vision with close parallels to the human visual system and eye movements.

Fractals have been used successfully in generating missing information in scene reconstruction. If this information is processed using coherent optical techniques, then it is important to understand the properties of the Fourier transforms of fractals. Since fractals with square or rectangular bases are representative of pixel structures found on most electrically adressed spatial light modulators, a square-based fractal has been generated and its far field diffraction pattern calculated. The properties of the resulting Fourier spectrum are discussed.

Barnsley and Hurd classify the fractal images into two families: iterated function system fractals (IFS fractals) and fractal transform fractals, or local iterated function system fractals (LIFS fractals). We will call IFS fractals, class 2 fractals and LIFS fractals, class 3 fractals. In this paper, we will unify these two approaches plus another family of fractals, the class 5 fractals. The basic idea is given as follows: a dynamical system can be represented by a digraph, the nodes in a digraph can be divided into two parts: transient states and persistent states. For bilevel images, a persistent node is a black pixel. A transient node is a white pixel. For images with more than two gray levels, a stochastic digraph is used. A transient node is a pixel with the intensity of 0. The intensity of a persistent node is determined by a relative frequency. In this way, the two families of fractals can be generated in a similar way. In this paper, we will first present a classification of dynamical systems and introduce the transformation based on digraphs, then we will unify the two approaches for fractal binary images. We will compare the decoding algorithms of the two families. Finally, we will generalize the discussion to continuous-tone images.

We consider an iterative map derived from the device equations for a silicon p⁺-n-n⁺ diode, which simulates a biological neuron. This map has been extended to a coupled neuron circuit consisting of two of these artificial neurons connected by a filter circuit, which could be used as a single channel of a parallel asynchronous processor. The extended map output is studied under different conditions to determine the effect of various parameters on the pulsing pattern. As the control parameter is increased, fixed points (both stable and unstable) as well as a limit cycle appear. On further increase, a Hopf bifurcation is seen causing the disappearance of the limit cycle. The increasing control parameter, which is related to a decrease in the bias applied to the circuit, also causes variation in the location of the fixed points. This variation could be important in applications to neural networks. The control parameter value at which the fixed point appear and the bifurcation occurs can be varied by changing the weightage of the filter circuit. The modeling outputs, are compared with the experimental outputs.

A computational theory and neural architecture for determining image affine flow and binocular affine disparity is presented. The computation of image affine flow is formulated as a system of linear equations, and the computation of binocular affine disparity is formulated as a dynamical system defined on the parameter space of a Lie subgroup of the 2D affine Lie group. The proposed neural architecture includes a set of neurons called the Lie-germs which function as the Lie-derivative operators, a set of simple cells with dynamical receptive fields, a set of intrinsic neurons that can affine transform the receptive fields of simple cells, and an analog circuit for determining affine parameters. The result of computer simulations of the proposed neural architecture for binocular affine disparity is also presented in this paper.

An active feature matching technique was developed, which incorporated both local and global information in the matching process, to achieve a global optimal goodness-of-match. First, an optimal snake was developed to reduce the 2D object of interest to a 1D feature string. This snake has the capability to extract accurate information about an object's corners that contains critical important discriminant information. High performance was achieved by dividing the energy optimization process into multiple stages that optimized both performance and speed of the snake. After the objects to be matched were reduced to two feature vector strings, dynamic feature matching (DFM) was used to match these strings. DFM matched the two feature strings in a global optimally way by using the Bellman optimality principle. An active image registration system was then developed using active feature matching to obtain a partial disparity map from which a full disparity map was estimated using regularization. This system was tested on a sequence of MR functional brain images. Results showed that the brain activation map obtained from registered images was significantly improved when compared to nonregistered images. Finally, an active image recognition system was implemented based on active feature matching. This system was applied to aircraft images and results showed that the active recognition system had superior distortion tolerance over the correlation based system and maintained good performance over a wide range of distortion. This tolerance to distortion was due to its 'active' nature. In other words, it, to some extent, mimicked human vision by dynamically adjusting the matching path so that the differences due to perspective distortion were minimized.

We present a range-finding method for determining surface topography by using time modulated structured light illumination. By illuminating a surface with a time-modulated light structure of several transmitters and triangulating relflected light onto a single point receiver, we are able to determine the surface height at a point by determining which light source intersects the surface at that point. In essence, the surface height acts to multiplex the projected signals onto the receiver. We increase the system resolution beyond the number of light sources by overlapping the image intensity of the light sources. The resulting signature allows for sub-pixel resolution. Time modulating the structurted light allows for demodulation with a high signal-to-noise ratio without the use of high intensity light sources thereby reducing system cost and complexity. This method of topographical analysis can be scaled to high end systems capable of real-time, high-resolution imaging. In addition to system geometry and resolution capacity, we discuss the advantages and disadvantages of various modulation and coding schemes that can be applied to this approach.

This research involved the set-up and demonstration of an optically addressed hologram writing experiment. The bulk of the research was devoted to determining the most efficient architecture and optical devices. Results indicate that the overall goal could be accomplished with the right spatial light modulator and the right geometry to allow the spot size needed for holographic recording.

Retinal pathology often results in visual field loss. Age related macular degeneration (ARMD) and Stargardt's disease (a congenital disease) typically affects the central macular region of the retina, resulting in visual field loss to the region of the retina with the highest resolution. Due to the central visual field losses, patients with ARMD and Stargardt's disease often experience difficulty in visual tasks, such as reading and facial recognition. NASA Johnson Space Center has developed an electronic remapper that can warp an image from one coordinate system to another on a television screen. The remapper can be used with patients with central visual field loss to redistribute information projecting off of the macular lesion (corresponding to the central scotoma) and onto the still functioning peripheral retina. The purpose of this research project was to investigate whether remapping of text around the central scotoma improved reading performance (increased reading rate) for patients with ARMD or Stargardt's disease. The subjects moved the text on the stage and read aloud random words of equal difficulty viewed on a closed circuit television screen. Reading speeds for normal and remapped text were obtained. Reading rates were determined for both free viewing and with stabilization of the position of the screen relative to the eye rotations. Results of these experiments are discussed.

In the transmission of remote images, a typical scenario may be one in which a user is searching through an image database which contains a large number of different images, the user may have to download (from a remote site) the complete image itself. Instead of downloading the complete image, it would save time and network bandwidth by first sending a much compressed version of the image (its identification image) that gives sufficient information to allow the user visually classify the image and then transmit the complete version of the desired image. In this work, we present a computationally very simple approach to create image identification for viewing clasification purpose. The new scheme compresses and transmits only the information in and around the edges of the image and a simple interpolation technique is used to construct the ID image from those compressed features at the user's computer. The technique described herein may be useful in telebrowsing image databases.

Rank and median filters (MFs) have become one of the standard techniques for suppressing noise of impulsive nature. A limitation of these filters is the high computational cost when implemented on general purpose computers. Several fast algorithms have been developed for specific types of signals. An alternative is computationally more efficient filters with MF-like performance. One proposed filter in this category with selectable trade-off between reduction of computation and MF-like performance is the block median filter (BMF). The BMF produces a median block at each window position, thus effectively reducing the total window steps and the computation. In this paper some statistical properties of the BMF and comparisons with the MF are presented. It is shown that for large window sized, the statistical properties of the BMF approach those of the MF.

Digital mathematical morphology (MM) and the distance transform (DT) have many points of intersection. The DT combines numerical features and object shapes. Usual application of mathematical morphology uses distance transforms based on the trivial city-block or chessboard metrics like digital representation of continuous ball. In this paper we concern structuring element as oriented neighborhood structures (ONSs) defined in the digital space; wider class of digital metrics and nonmetric functions defined on these NSs. We show that basic morphological operations may be performed by a thresholding of results of local propagations based on this class of digital functions. Such approach simplifies a description of some structuring elements and some operations of image analysis.

A neighborhood operation binary image algebra (NOBIA) which has only one basic operation is presented and developed. The basic operation of NOBIA is a convolusion followed by a nonlinear filtering function and then an intersection operation. The parallel architecture of an optical neighborhood operation digital image processor is designed to efficiently perform the parallel morphological image processing algorithm provided by NOBIA, and the optical implementation hardware of the processor is discussed. Using an incoherent optical convoluter as a 3D free space interconnection device, a smart LCLV as an optical nonlinear device, and a simple optical circuit for intersection operation, an optical hardware is constituted and used to realize the basic operation of NOBIA experimentally. Any image transformation can be performed with this hardware by executing the basic operation repeatedly according to proper iterative programming.

Texture is an important feature of images and it has been widely used for image analysis. Peleg has proposed that the fractal signature model can be used for classifying textured images. Fractal signature is the slope of measured area of a gray-level surface with changing resolution. To classify the textured image he suggested to compare fractal signatures in terms of a weighted mean squared error measure. In this paper, we propose that the area distribution with several similarity measures rather than its slope (fractal signature) can also be used for the same purpose. The area distribution of image is fed directly to the input of the classifier. Depending on how to design the structure of the classifier and what similarity measure to choose, the classification ratio is much different. We compare the methods, and show that as high as 83-93% correct classification ratio can be achieved. The texture pictures in this paper are taken from Brodatz. We find the best performance can be achieved when the area distribution rather than fractal signature is used with the two-layer perceptron classifier trained with the back-propagation algorithm.

A new DSP-based neural simulating computer architecture and its ANN-based assignment method for parallel distributed processing are proposed. The hardware of the proposed neural simulating computer can be reconfigured in terms of a variety of research interests and requirements of pattern recognition. The software programming environment utilizes an intelligent compiler to perform static task assignment in both the cases of single-task muliprocessor and multitask processor. An improved Hopfield neural network which can converge to global optical solution is employed by the compiler to map different tasks or neurons to their corresponding real processors. An approach of introducing hidden layer to increase the computation ability of the neural simulating computer is also developed. Finally, a proof is given which shows that the use of improved Hopfield algorithm and the modification to network structure doesn't change the intrinsic properties of the original network.

In this paper, we describe a computationally efficient 3D object surface matching algorithm. In the proposed method, object and model surfaces are scaled to be in a unit cube in the 3D space. They are then sliced along the magnitude axis and the resultant object and model surface cross sections are represented in binary image format. The cross sections' centroids of an unknown object and the models of different shapes are computed in their respective binary images. The resultant cross-sections are translated to the origin of the spatial plane using the centroids. Major and minor axes of the plane cross sections are aligned with the coordinate axes of the spatial plane. Matching of the aligned cross sections is done in the direction of the gradient of the cross section distance between the object boundary points and the corresponding points in the model cross section boundary. The shape deformation distances measured in different cross sections are averaged and the minimum average shape deformation distance is used to identify the model best matching to the object of unknown classification.

In this paper we present the relational graph description of natural color scenes for model- based matching using relational distance measurement. The uniformly colored object areas and the textured surfaces of natural scenes are extracted using color clustering and linear discriminant. The extracted object regions are refined in the spatial plane to eliminate the fine grain segmentation results. The refined segments are then represented using an adjacency relation graph. Scene model is formed by means of 3D to 2D constraints and adjacency relations. The relational-distance measure is used for matching the relational graphs of the input scene and the respective image. Experiments are conducted on imperfect color images of outdoor scenes involving complex shaped objects and irregular textures. The algorithm has produced relatively simple relational graph representation of the input scenes and accurate relational-distance-based matching results.

This paper concerns geometric transform invariant texture analysis, and in particular texture analysis algorithms whose performance is not affected by changes in scale, rotation angle, or perspective projection. The importance of geometric transform invariance is first discussed, and an overview of the existing scale and rotation invariant texture analysis approaches is then presented. The importance of perspective invariance is emphasized and current work on the subject is summarized. Directions for possible future research are suggested.

An adaptive cosine tranform coding algorithm is proposed for data compression. The algorithm employs cosine transform and marginal analysis technique to convert data from time domain to frequency domain and achieve bit allocation for the quantization of the signal in its frequency domain. Compared with the conventional cosine transform coding which employs a logarithm approach to bit allocation, this algorithm yields lower mean squared quantization errors. The simulation result is presented here for waveform signal coding but, with a 2D cosine transform, the algorithm is applicable also to image signals.

We investigate two specific visual motion behaviors that are appropriate for mobile robot applications because of their simplicity and low computational load. The first behavior determines the distance to a flat wall to the side of the moving robot. Such a measurement is essential for a wall-following behavior. The second behavior determines the projected amount of time until the robot collides with a stationary object assuming that the robot continues to move at a constant velocity. This time-to-contact estimate relies on the expansion of the optical flow field and can act as a collision-warning sensor to warn of impending obstacles. The numberical estimates given by both behaviors are expressed as simple ratios of easily measured visual quantities and do not require any elaborate calibration procedures. These behaviors use a 1D patch-wise correlation technique that was developed by Poggio and Ancona. We have further simplified the computation and show results on true image sequences.

In this paper, we propose a new model of spatial frequency column (SFC) in primary vision pathway based on multiresolution quasi-wavelet frequency plane sampling. The proposed model uses a set of parameter variable filters to sample frequency plane to distinguish different frequency components. The design of these filters is based on following criteria: 1) Channel filters should be well localized in both time and frequency domains. According to uncertainty principle, the Gaussian function is a best candidate. 2) The overlap between filters should be as less as possible to facilitate delimitation of distinctive frequency channels. 3) The filter bandwidth and its center frequency is a constant under same sampling resolution since vision system is more sensitive to higher frequency components. 4) The sampling density in frequency plane can be adaptively controlled, so that the obtained channels are not redundant and the details of signal are not lost, which means the sampling should be of multiresolution. When the insufficiency of sampling density at a point is detected, the sampling algorithm can adaptively increase the sampling density in the vicinity of that point through raising the resolution parameter. The sampling algorithm can be simply described as follows: 1) Define a lower resolution. 2) Use above criteria to yield N filters. 3) Convolute the image with the filters to produce N outputs. 4) Detect the sufficiency of sampling density at each filter center, if the sampling is too sparse then increase the resolution at that point. This is a recursive procedure. 5) At each pixel, take the maximum value of its frequency distribution set as its estimation value. Some experimental results show that the proposed model can partly imitate the SFC in the primary vision pathway.

Both in 1D (signal analysis) and 2D (image processing), the wavelet transform (WT) has become by now a standard tool. Although the discrete version, based on multiresolution analysis, is probably better known, the continous WT (CWT) plays a crucial role for the detection and analysis of particular features in a signal, and we will focus here on the latter. In 2D however, one faces a practical problem. Indeed, the full parameter space of the wavelet transform of an image is 4D. It yields a representation of the image in position parameters (range and perception angle), as well as scale and anisotropy angle. The real challenge is to compute and visualize the full continuous wavelet transform in all four variables--obviously a demanding task. Thus, in order to obtain a manageable tool, some of the variables must be frozen. In other words, one must limit oneself to sections of the parameter space, usually 2D or 3D. For 2D sections, two variables are fixed and the transform is viewed as a function of the two remaing ones, and similarly for 3D sections. Among the six possible 2D sections, two play a privileged role. They yield respectively the position representation, which is the standard one, and the scale-angle representation, which has been proposed and studied systematically by two of us in a number of works. In this paper we will review these results and investigate the four remaining 2D representations. We will also make some comments on possible applications of 3D sections. The most spectacular property of the CWT is its ability at detecting discontinuities in a signal. In an image, this means in particular the sharp boundary between two regions of different luminosity, that is, a contour or an edge. Even more prominent in the transform are the corners of a given contour, for instance the contour of a letter. In a second part, we will exploit this property of the CWT and describe how one may design an algorithm for automatic character recognition (here we obviously work in the position--range-perception angle--representation). Several examples will be exhibited, illustrating in particluar the robustness of the method in the presence of noise.

Novel integral videosensor and its possible application for real-time operation in machine vision, robotics, and visual communication are discussed. This sensor is a solid-state photoelectric structure with memory (PESM), which consists of semiconductor, insulator, and metal layers, wherein visual information is registered, stored, and processed as 2D charge and potential patterns. So, the PESM operates as a multifunctional device utilizing massively parallel processing due to internal interactions between charge and potential patters, induced electric fields, and incident illumination. That is why it may be effectively applied to the real- time visual processing optoelectronic systems. In this report we would like to propose some ideas and methods of the PESM applications in intellectual machine vision system. The basic image operations of the PESM, namely subtraction and correlation, allows one to solve in real time the main problems concerned with the selection of a required object with subsequent tracking, pointing, and guidance. It is essential that in these and similiar PESM applications for the optoelectronic system the usage of additional electronics is minimized, involving its substantially lower prime cost compared with the now existing systems. As we suppose, the possible applications of the PESM as an intellectual 'supervideosensor' that combines a high quality imager and a high-capacity processor, may be widely extended.

This paper focuses on the design of segmentation based on textural features extraction. This is an example of a transposition between biological and visual phenomena used in order to characterize natural image understanding. This is also an illustration of a more general appraoch to IP knowledge representation based on a methodology dedicated to the formalization of concrete and abstract models for image processing applications. It proposes an ontology which includes conceptual specifications borrowed from mathematics and physical and biological axiomatics which give concrete and more natural sense to our IP models. This provides a set of elementary definitions which can be used for the expression of concrete models such as image segmentation or pattern detection. In the case of texture we would like to formalize grey level behavior through processings based on multiple window analysis (spectral and morphological criteria: grey level and compacity). In this framework, the evolutionary models studied are issued from biological modeling of migration and mutation. Our illustration is relevant to multispectral segmentation where the homogeneity criterion has been modelized by a grey level evolution function based on exponentiation.

In diverse technical units, there are widely represented surfaces of complex figures for which strict requirements in respect to control accuracy and efficiency, are specified. Currently these surfaces are as a rule inspected using universal or specific measurement means characterized by discrete measuring of the location of points or separate profiles of surfaces under control. A task at which the present work is aimed comprises development of a technique and of installation to provide simultaneous information about the whole surface controlled, substantial heightening of the measurement accuracy, high rate of inspection, and rigidity to systems mechanical oscillations.

In a number of applied areas, a problem in respect of automatical joining (co-sewing) of image frames, being recorded element-by-element, does arise. To that, there emerges a problem of selection both of joining related algorithm and effective prefiltration. Voluminous literature is dedicated to a solution of these problems. The present work proposes a phase method using Walsh's 2D functions.