An approach to automatic threshold selection based on information gain, is proposed. The probability distribution of gray levels can be utilized to compute the information content of the image using entropic measures. An objective function representing the information gain provided by the threshold gray level with respect to the other gray levels is defined. The gray level value that maximizes the discriminant function is selected as the threshold value for the input gray level image to provide an output binary image.

When an object moves relative to a sensor, the motion of its image in the image plan produces an image flow field. The calculation of image flow which involves obtaining a velocity vector at each pixel is considered very important in the area of robot vision. This paper introduces a method for the calculation of optical flow generated by the motion of planar objects under orthographic projection. The proposed method consists of two phases. In the first phase an interpolation technique is introduced to obtain the correspondence of dense feature points. In the second phase a least squares method is used to estimate the optical flow parameters from the corresponding point pairs obtained in the first phase.

This paper presents two new approaches for optimum detection of edges and contours in a gray level image of a scene with a constant background. The first algorithm determines the optimum tolerance for edge tracking that maximizes the number of edge pixels in each object in the image, while the second algorithm utilizes an iterative thresholding scheme in conjunction with a golden-section optimization technique to determine the optimum threshold that maximizes the number of contours and the number of pixels in the contours.

Basic image processing operations such as thresholding are often difficult to use in industrial settings because of nonuniformities in scene illumination. However, these illumination variations can be reduced or eliminated entirely if appropriate image processing techniques are applied prior to level-sensitive processing. Useful procedures for reducing the effects of illumination variations are covered in this paper including background subtraction, high-pass filtering and the use of homomorphic filters. Hardware considerations for performing these operations in real-time systems are also discussed.

An image may have different local statistics, or different local properties of detailed content. According to the information theory, a fixed bit rate coding scheme will result in more distortion in higher-detail (or active) areas than in lower-detail (or non-active) areas. Consequently, an edge degradation occurs. In this paper, an adaptive vector quantization with least squares approximation method is proposed to address this problem. The scheme is based upon a variable block size. The basic idea is to partition the image blocks into two parts; high-detail and low-detail or active and non-active, according to the activity index measurement. The active blocks are further divided into smaller blocks. The values of pixels in the non-active block are approximated by a smaller number of samples with least spuares method. The two sets which are designed to have same dimensionality are used to form a vector set and vector quantized. Therefore, the non-active blocks can be coded with the same codebook as for the high-detailed subblocks, but less bits are needed. Experimental results show that the quality of reconstructed images is improved in this algorithm over non-adaptive VQ because the bits saved in two low-detailed areas are allocated to the high-detailed area where the edge could be found.

Quantitative measurements of surface variations are essential to computer inspection of surface quality. We propose a new surface variation measurement algorithm which computes a surface roughness map via the surface normals. The sampled surface data are first approximated by bivariate polynomial functions under least squares criterion. Surface normals are then estimated from the polynomial coefficients. A roughness metric is determined for each surface point in terms of the cumulative angular differences between pairwise normal vectors in the neighborhood. Experimental results show that the established roughness metric can distinguish subtle details of minor surface variations. The computation required in this algorithm are convolution-based operations. Therefore, the method can be readily implemented in a parallel computing environment for real-time applications.

Recently we have proposed a range image segmentation method based on the equidistance contour map extracted from a range image. An equidistance contour of a range image is formed by pixels on the range image whose corresponding scene points are all at a same specified distance from the sensor. We have observed that in different ways the contours reflect the existence of object surface edges, the geometries of object surfaces, and the orientations of these surfaces in the 3-D space. In this paper we present a method which uses the equidistance contours to further characterize the regions obtained from segmentation. This is an integral part of an object recognition process. The meaning of characterization is to find the geometries of the corresponding 3-D object surfaces of the regions from the contours. If a surface has nice analytic form such as planar, spherical, cylindrical, or conical, we determine not only its type but also the values of the parameters which describe its geometry.

The gray level geographical structure (GLGS) is a simple method to represent the local intensity variation of an image in symbolic description. This representation can be used in higher level image processing in subsequent steps. The advent of VLSI microelectronic technology has led to the idea of implementing the GLGS directly in hardware. A two dimensional pipelined systolic pixel classification array is proposed in this paper. In the design, each pair of processing elements processes the data in a pipelined fashion and the data in each pair of processing elements is processed in a parallel fashion to further enhance the system performance.

Shadow Vision: determining the dimension and position of an object by measuring its shadow cast from a collimated (parallel) light source. This alternative technology for measurement and inspection tasks such as edge detection, dimension measurement, and web inspection is presented. An explanation of shadow vision and differences from conventional systems is described with application examples. A modular collimated light source design is discussed which produces a beam approximately one inch high by seven inches wide. The beam can be extended to virtually any width. Discussion of recent developments in amorphous materials shows how a linear array photosensor is produced on a steel substrate having densities higher than in previous technologies. A sensor unit has been developed using these photodiodes with densities of 30 to 40 photodiodes per inch. It too can be made virtually any width.

Major limitations to the successful application of optical pattern recognition systems have usually been the memory requirements necessary for realistic tasks and the implementation of such optical memory techniques. Here, we have considered the possibility of generating the NXN array of filters by using real-time computer generated holography where the Fourier transforms of the NXN reference image are produced in the computer. The NXN array of Fourier transform holograms are converted to phase-only encoded filters by utilizing the phase function of the Fourier transform elements of the array. The phase-only encoded NXN array is written onto a spatial light modulator for pattern recognition applications. Thus, a phase-only encoded correlator with high storage capacity is produced. An important feature of the proposed technique is the ability to update or change each element of the NXN filter array in real-time independent of the other members of the filter set. This feature does not exist in the previous large memory correlation techniques since the filters were stored on film and to change a member of the array required a new synthesis of the entire array. We shall study the performance of the proposed binary capacity correlator by determining the peak to sidelobe ratio and the bandwidth of the resulting correlation signals. The effects of phase-encoding and the finite space-bandwidth product of each element of the array will be studied. The effects of overlapping terms at the filter plane contributing to cross talk will also be investigated. Both binary phase-only encoding and continuous phase-only encoding are investigated and the results are compared to the multiplexed classical matched filters.

A multiobject shift-invariant pattern recognition system using code division multiplexed binary phase-only correlation is presented. The system computes the binary correlation between an input pattern and a generalized set of pattern functions. This technique uses a filter which consists of a set of binary phase-only code division multiplexed reference pattern functions. There are many advantages in binarizing the filter function. Binary spatial light modulators (SLMs) have been developed that work well in a binary phase-only mode and can be used to synthesize the spatial filters of this type. Binarization also permits the recording of filters of images with larger samples on currently available binary SLMs that have a limited number of pixels. The functions in the reference set may correspond to either different objects or different variations of the object under study. Computer simulations of the correlator are used to study the performance of the pattern recognition system. The correlation signal-to-noise ratio (SNR) and the ratio of the correlation peak intensity to the maximum correlation sidelobe intensity are evaluated as the criteria for the system performance.

A high-performance, high-capacity associative neural memory (ANM) architecture is proposed. The proposed ANM architecture is based on a cascade of two-levels of fully interconnected layers of binary neurons. Feedback is used to connect the output of the second neural layer to the input of the first layer in order to increase network convergence rate and retrieval accuracy. The proposed ANM training (recording) is accomplished by the use of a very efficient newly-developed recording technique. The combination of the above powerful recording technique and the two-level with feedback ANM architecture gives rise to a high-performance network for pattern recognition applications. The proposed architecture allows for simultaneous autoassociative and heteroassociative memory operation which implies both pattern reconstruction and pattern classification capabilities. Finally, the highly parallel and distributed architecture of the above ANM can greatly benefit from the intrinsic parallelism and high interconnection capacity offered by optical systems.

In this paper, we compare the performance of the bipolar joint transform correlator to the continuous phase-only matched filter correlator, the binary phase-only matched filter correlator, and the classical correlator. The pattern recognition systems will be compared in the areas of correlation peak, autocorrelation peak to sidelobe ratio, autocorrelation bandwidth, and discrimination sensitivity.

A wedge-ring Fourier feature extractor is enhanced with Fourier plane intensity spatial filtering to provide image-domain, discrimination feature displays. A brief review is given of the optical Fourier feature extractor architecture and its practical industrial inspection applications. A previous limitation of Fourier feature extractors has been the lack of display to the end-user of discrimination features in the space- or image-domain. A solution is presented consisting of a specialized 16-wedge and 16-ring Fourier plane intensity filter set. From Fourier feature vector data produced by the feature extractor, Fisher discriminant ratios determine the Fourier plane spatial filters to be used. Thus a dynamic architecture with feedback is possible, allowing the end-user to adjust front-end video image acquisition and/or object presentation to maximize the Fourier feature space novelty of defects to be detected. Computer software generated spatially filtered images are presented. An approach for implementing dedicated hardware is suggested, utilizing optical Fourier plane multiplication to achieve real-time convolution.

A joint Fourier transform image correlator that uses a binary spatial light modulator at the Fourier plane is presented. In this system, the Fourier transforms interference intensity is thresholded to only two values so that a binary spatial light modulator can be used to read-out the interference intensity. The performance of the correlator introduced here is compared to the classical joint transform correlator in the following areas: correlation peak, correlation peak to sidelobe ratio, correlation bandwidth, and discrimination. We show that the bipolar Fourier transforms interference intensity can result in a higher correlation intensity and a better defined correlation spot than the classical joint transform correlator. Computer simulation is used to test the correlators and the three dimensional plots of the autocorrelation and cross-correlation functions are presented and discussed.