Engineers at the Rome Air Development Center (RADC) are investigating novel approaches to automating the exploitation of digital imagery obtained from reconnaissance missions with the ultimate goal of developing an automated imagery exploitation system. Incorporating many computer vision and image understanding algorithms such a system would automatically detect segment classify and identify all ground targets in a specified area of interest. Development challenges and some proposed solutions are discussed in an attempt to stimulate further research.

The food industry offers many automated inspection challenges that are not readily solved using commercial industrial vision systems. A discussion of these challenges includes a package inspection involving a variable " good" package color analysis of a baked product high speed verification of a coating process and a high speed printed web inspection. Solutions for these problems are solicited.

This paper describes a way of applying statistical estimation techniques to the model matching problem of computer vision. A detailed metrical object model is assumed. A simple MAP model matching method is described that captures important aspects ofrecognition in controlled situations. A probabilistic model of image features is combined with a simple prior on the pose and feature interpretations to yield a mixed objective function. Extremizing the objective function yields an optimal matching between model and image features. Good models of feature uncertainty allows for robustness with respect to inaccuracy in feature detection. Additionally the relative likelihood of a feature arising from the object or the background can be evaluated in a rational way. The parameters that appear in the probabilistic models may easily be derived from images in the application domain. The objective function takes a particularly simple form when feature deviations are modeled by Normal densities and the projection model is linear. Several linear projection and feature models are discussed. Evidence is provided to show that Normal feature deviation models can be appropriate for computer vision matching problems. Relation to other work and possible extensions and application areas are discussed.

To avoid the excessive computation of testing all combinations of feature matches groups of model features can be arranged in an index space offline (hashed). Ideally each image group should index into the space and find only those model groups that could have formed that image group. We prove an unexpected tight lower bound on the space required for such an indexing scheme for point features 3D models and 2D images and consider some implementation issues.

Individual facial features such as the eyes or nose may not be as important to human face recognition as the overall pattern capturing a more holistic encoding of the face. This paper describes " face space" a subspace of the space of all possible images which can be described as linear combinations of a small number of characteristic face-like images. The construction of face space and its use in the detection and identification of faces is explained in the context of a working face recognition system. The effects of illumination changes scale orientation and the image background are discussed.

A system for identifying complex space objects from a sequence of wideband radar images is presented in this paper. The system is referred to as the Complex Space Object Recognition System (CSORS) and uses a data-driven approach to object recognition. The input to the system is a time sequence of range/Doppler radar images of an object in orbit as the object passes overhead. The system first processes the individual images to improve the signal-to-noise-ratio and then further processes each image to derive a set of features. The sequence of feature sets for each pass of the object is then processed to produce a three-dimensional wireframe of the object. Finally, a symbolic model representing the object is generated from the wireframe. This model is matched against a library of models and appended to the library if a match is not found.

A learning procedure is described for the recognition of 3d industrial objects from 2d images. It is assumed that the objects are solid and have well defmed edges and that viewpoint and lightning are well defined but that there is no information available on the orientation distribution of future objects to be classified. The presented learning procedure covers all orientations by an initial sampling detects gaps and deletes superfluous orientations. An example is presented.

The pattern spectrum describes the shape and size of structures in an n-dimensional signal. Measurement of the pattern spectrum is based on morphological operations which use a variety of structuring elements to filter a signal at multiple spatial scales. This paper reports on the use of pattern spectra in the grayscale domain for classifying different textures and in the binary domain for object recognition. The advantage of morphological image processing is that it is based on highly parallel primitive operations which are amenable to large-scale implementation in real-time signal processing hardware.

An application of dynamic programming based matching to vehicle model identification is proposed. An edge map is first generated and various pre-processing stages are carried out culminating in the use of a 2-D Hough transform to generate a cued region containing a putative car. A dynamic programming search (Edge List Search) is then carried out to identify the type of car e. g. Citroen Lada mini van etc. within the located area. Experimental results are given.

This paper presents a flexible and highly-reliable gray-level vision system based on multiple cell-feature descriptions using only three basic operation modules: extended convolution radially traversing probing and histogram compression. The generalized Hough transform is introduced as a universal method for object model matching. Model learning is automatically performed by acquiring image samples while rotating each object. A prototype system demonstrates successful recognition of mechanical parts.

This paper describes an algorithm which recovers 6 degrees of freedom ( that is rotational and translational matrices R and T ) of polyhedral objects using L-corners in a monocular image. First the gradients of the cirtain object''s faces are determined using the constraints derived from angles of L-corners. R accompanied by the object is then determined. Second a constraint concerning T is derived from a junction point of a Lcorner. After that T is determined using a set of constraints derived from the L-corners that contributed to the determination of R. By assembling only those L-corners the cirtain object is recognized.

The classification of complicated 2-D shapes such as cracks fractures and the abnormal patterns found on the surfaces of materials is very important in automated surface inspection. As a method of facilitating the classification of these shapes we report a strategy of extracting the hierarchical structures from these complex patterns using branch cutting boundary analysis and others. Experimental results are given with random statistical patterns and with the images taken from the surfaces of some materials.

Perceptual grouping of image features are groups of 2D image features which possess properties that reflect their 3D strucwres they are useful for matching 2D image features to 3D models. In this paper perceptual grouping experiments based on cyclically connected loops and partially out-of-view loops are performed. Results can be applied to separate individual objects in the scene filtering out noise and pruning the search space for matching the image features to model features.

This paper outlines a specific method of cooperative edge linking, where difference (edge points) and homogeneous (boundaries of regions) informations are simultaneously available. In only one scanning of the image, both classical steps of "edge closing (in a small distance)" and "edge linking" are gathered. Complete chains, corresponding to strong edges, are first extracted, knowing the local edge configuration around each edge point. Then, they are brocken, in order to find junction points. Such junctions are searched only at both ends of chains, on the edge image.

Computer diagnosis of the vascular diseases treatable by laser burning (photocoagulation) are investigated. Specific concern is with identification of the region and its distance from the center of the foveal avascular zone. The parametrization of the lesion boundary and location and measurement of distance from the lesion to foveal center is accomplished with an algorithm which builds the boundary as points on spokes emanating from a chosen interior point one point on each spoke.

The problem of palletizing (stacking on a pallet) randomly arriving mixed size and content parcels is an important task in most distribution warehouses. Today this task requires human interaction for a solution however recently several attempts have been made to automate the solution. The purpose of this paper is to present an overview of the problem an expert system approach and an estimate of the key subproblems which have been identified which are necessary for a solution. The concepts of space filling and emptying as encountered in warehousing are briefly described. Also brief descriptions of two generations of a robotic system for mixed parcel palletizing are presented. The results with these test systems indicate that automatic parcel handling at speeds comparable to humans is feasible however further work is required to obtain a robust solution.

A system which links a computer solids modeler to an active stereo imaging inspection station is used to inspect automobile glass. 1 Research concerning the method of generating projection patterns for circular contours is described. A prototype system consisting of a coniputer graphics workstation a projector and a machine vision system was assembled and the system''s performance is discussed.

This paper considers the control of two robotic manipulators jointly graspingwith no slippagea rigid object. Two conflicting control problems are addressed: Adaptively apportioning the object load between the manipulators and application of commanded interactive forces (or bias forces) to the object. By controlling the interactive forces damage to the manipulators and/or the object can be prevented. Additionally a desired tension torsion or cornpression can be applied to the object. Adaptive load apportioning (sharing) increases the carrying capacity of the system by using possible mechanical advantage. The Hopfleld net is used to minimize a quadratic energy function in the joint torques. The results are the optimal joint torques required to drive the load and supply the commanded bias force. Simulations are presented that show effective control of interactive forces while sharing the load in an optimal fashion.

Successful integration of robots into a manufacturing system is a crucial step in achieivein an efficient automated production system. However different robots use different languages. Io solve this problem standardization of robotic control languages is desirable. A standardized robotic control language UNIAN (UNIversal LANguage) is developed to control two different miniature robotic mampulators. UNILAN is designed as a motionbased language and is on a level close to the object level.

Vision systems will play an increasing role in inteffigent material handling systems. This paper discusses two hardware principles which make possible a host of cost-effective applications--integrated vision systems and the use of retroreflective materials. Described are (1) the design cost and performance characteristics of integrated systems those with the microcomputer array detector and illumination as part of a single circuit (2) the impact of using retroreflective materials those with apparent brightness of more than 1000 times that of diffuse white surfaces and (3) some specific applications: AGV guidance part handling and AS/RS control.

A new approach for representing 3-d objects by 3-d array grammar is introduced. The concept of universal array grammar is proposed for 3-d object representation. it uses parallelism for pattern generation. Many interesting 3-d objecis can indeed he represented by this universal 3-d array grammar which paves a ground for 3-d object recognition. description and understanding. Keywords: 3-d array grammars. universal array grammars pattern generation. object representation parallel generation I.

The complexity ofcomputing 3D workspaces forjoint limited redundant manipulators is examined. Different types of reachable workspace volume and reachable workspace boundary problems are defined. Each type of volume problem is at least as hard as its corresponding type of boundary problem and each problem type is at least NP hard. New efficient and adaptive workspace point computation techniques are proposed (e. g. based on nonlinear programming) after complexity analyses of the corresponding decision problems.

A new approach is introduced in this paper to deal with the problems of real-time machine vision and pattern recognition for robotic manipulations. This approach emphasizes three directions: (1) the developed algorithm has to be compact enough for embedded intelligent control implementation (2) the computational scheme should be highly efficient for on-line robot reasoning and manipulations and (3) the resulting system has to be sufficiently flexible to accommodate various working environments and to cope with some system shortcomings. The vertical integration of related vision hardware image analysis software and analytical techniques (e. g. fuzzy logic and neural networks) together with the novel algorithms for robot eye-brain-hand coordination constitutes a unique robot vision system. The potential of more extensive hardware implementation is discussed and a wider spectrum of applications of the proposed robot vision system is envisioned.

Arranging the photosensitive elements of an imaging sensor in a log-polar grid automatically samples an image in a logpolar space. The Retina project is a chip with such a spatially varying layout that can produce the advantages of image processing in the new space at real-time speeds. The actual chip is a small part of a complete imaging system. The system is part of a class of imagers called foveal sensors and these sensors have distinct and significant computational savings over conventional imagers as many as 3-10 orders of magnitude improvement in processing time and memory. The design maintains a large region of high-resolution data although it is still only a fraction of the total photosensitive area.

In previous work by ourselves and others it has been argued that the best algorithm for fusing multiple fixations of the same scene is to simply choose the maximum resolution information available at each position in the reconstructed scene. We can show under simple assumptions of pixel distributions that this algorithm is indeed optimal in a least-squared-error sense. In the presence of noise however optimality no longer holds a certain degree of averaging of lower resolution information is required to obtain optimal reconstruction. We also present empirical demonstrations of fused images using maximum-resolution and averaged blending techniqes and we illustrate the effects of noise on the quality of reconstruction.

This paper describes an ongoing effort to produce hardware that can perform arbitrary two dimensional image warping at a cost that is reasonable. This paper describes hardware that has been developed for the VNE bus. The hardware has been specifically designed to operate with the DATACUBE family of image processing boards. Following the description of this hardware is discussion on specific application research currently utilizing this technology. I.

In image compression for subsequent transmission over a data channel it is often desirable to discard nonessential information before further coding for compression is performed. In this paper polarlog coordinate mapping is used to reduce the input image in a way that preserves high resolution in the field of interest while collapsing information at the field periphery. This reduced image is then compressed by using runlength encoding thus allowing the compressed data to be transmitted over a narrower bandwidth channel. Simulation results for compression using polarlog as a preprocessing stage are presented. Other possible preprocessing stages as well as compression stages are also discussed.

A robot is a powerful production tool, however, an "intelligent robot" with flexible control is often needed to acomplish complex tasks and to adapt to changes in its working enviroment. This paper describes the integration of machine vision with a robot to accomplish this. The system integrates an IBM 7535 industrial 7535 Industrial Robot, with four degrees of freedom, with its own programming language, AML/Entry, and a Vision System from Intelledex Inc. The IBM has its own programming language AML/ENTRY, as does the Intelledex vision system, Vision BASIC. The major objective of the project was to interface the controller of the IBM 7535 robot with host processor (a compatible IBM AT) to a vision system. In an operational scenario, the vision system is tasked with the responsabilities of recognizing an object and determining its location and orientation. Based on this information, the robot is able to pick up the object and transfer it to a pre-defined destination. The paper describes how the interfaces were developed.

This paper presents a method for verification of ANSI form tolerances using computer vision with structured light. A sampling plan is developed to determine the required resolution of the system in order to inspect the part to a specified tolerance. A statistical method based on the measurement error is used for verification.

In Greiner'' formalisms for an analogical learning system were introduced. The method of analogical learning which resulted from these formalisms depended upon human intervention in the form of " analogical hints" which related objects for which analogical inferences could be drawn. In the work introduced here the formalisms have been altered in a way that permits the " surmising" of the " analogical hints. " In particular a general formula for analogical learning is introduced. With respect to some object preliminary work is introduced which will result in a method for deriving unknown elements in the general formula from the known elements. Unknown/known elements in the general formula will include knowledge concerning the components which comprise an object knowledge concerning how to perform certain tasks using an object and how the object relates (from an analogical viewpoint) to other objects (i. e. the analogical hints). 1 . 0 OVERVIEW Learning is a key core technology in the development of intelligent robotic systems. In the past most efforts toward the development of learning systems to support robotic applications have centered upon the navigation and task planning issues. These approaches have combined the use of spatial/geometric reasoning with learning algorithms for example simulated annealing. The work reported herein demonstrates the first steps toward the development of a generalized learning system and is based upon work reported to NASA2. Based on these results we

In this paper we describe a system for monitoring activities in aerial images. The system uses many knowledge sources and databases to guide the interpretation process. It has an innovative control scheme for managing the execution of the tasks utilizing of its knowledge sources and updating of its many databases. As a result the system not only is aware of the existence and usefulness of all knowledge present in its vanous databases it also knows how to use that knowledge and the processing modules in cooperation to achieve a common goal of detecting events of interests from a set of aerial photographs. Also to overcome the difficulties of the object recognition task a model-based technique is used to detect predict verify and recognize the objects of interests in the scene.

Fuzzyc-Efflpsoidal Shell (FCES) algorithm that utilizes hyper-ellipsoidal-shells as cluster prototypes is proposed. FCES is a generalization of the Fuzzy Shell Clustering (FSC) algorithm. The generalization is achieved by allowing the distances measured through a norm inducing matrix that is symmetric positive definite. In case offixed known norms the extension of FcS to FCS is straightforward. Two different strategies are recommended when the norm is unknown. The first strategy considers use of non-linear least-squared fit approach with fuzzy memberships as weights. The second approach considers norm inducing matrix as a variable of optimization thus making FCES an adaptive norm type algorithm. An adaptive norm theorem is presented. The results of first approach is used to detect ellipses having unequal sizes and orientations in two-dimensional data-sets. Non-linear equations of the FCES algorithm are more complex than those of the FSC algorithm. Numerical issues related to both the FCES algorithm and the FSC algorithm are discussed.

A recognition system which represents object categories by properties which can he deduced by analysis of 3-D shape has been implemented and tested using the category " chair" for case study. Functional description is used to recognize classes and identify subclasses of known categories of objects even if the specific object has never been encountered previously. Interpretation of the functionality of an object is accomplished through qualitative reasoning about its shape. This is to our knowledge the first implemented system to explore the use of purely function-based representation (that is no geometric or structural object model) to recognize 3-D objects. During the recognition process evidence is gathered as to how well the functional requirements are met by the Input structure. This paper is concerned with choosing the type of operators that will be used in the combination or accrual of the functional evidence. Three pairs of conjunctive and disjunctive operators aie evaluated. Each pair is uSe(i in the recognition process of the 100+ test objects. The results of all tests run are compared and diffeieiices are discussed.

In this paper, the ability of the retina to detect motion in the retinal peripheral visual field is emulated. The process of emulation utilizes a special receptive field defined in the spatial and temporal domains. The simulation studies on this emulated peripheral visual field show that it is possible to acquire certain robust abilities in artificial systems which are inherent in their corresponding biological processes. Abilities such as noise suppression were evident in the emulated motion detection system presented in this paper.

Attributed Fuzzy Tournaments (AFT''s) are a special type of attributed fuzzy graphs which are useful to represent uncertainties inherent to many realworld problems. A new algorithm that finds the best fuzzy matching configuration between components of two Attributed Fuzzy Transitive Tournaments (AFTT''s) is proposed. The best fuzzy matching between two AFTT''s is the matching configuration between components of both AFTT''s such that the overall distance measure between two AFTT''s possesses the minimum value. Useful applications of the proposed algorithm can be found in scene matching where the nodes of an AFT represent the objects in the scene and the arcs represent the relationships among the objects. Uncertainties of the image are represented via fuzzy membership values associated with the nodes and arcs. An examples showing the usefulness of the algorithm in pattern matching is shown through image analysis. I.

We look at structure of the fuzzy logic controller from the point of view of a function relating input to output. This view in turn can be generalized to that of looking at the fuzzy logic controller as a mathematical relationship. In this view the problem of fmding fuzzy logic controller outputs becomes equivalent to solving mathematical relationships. Using this motivation we suggest an alternative approach to the calculation of the output of fuzzy logic controllers. This approach is based upon the use of level sets and is called the level set method. This method provides a very efficient way of calculating fuzzy logic controller outputs that is easily implemented in hardware.

The fuzzy integral has been shown to be a very effective tool in the fusion of information for decision making. It combines objective evidence with possibly subjective information concerning the worth of this evidence as incorporated in a fuzzy measure. In many real applications the degree of importance of a source of information should be computed from a set of training data instead of being supplied subjectively. In this paper a mathematical justifiable approach for the automatic generation of fuzzy density values (the basis of the fuzzy measure) from training data is presented.

Advanced real-time digital controls for complex plants or processes will use a model (an " Observer" ) which predicts the values for sensor readings expected from the actual plant these vote as alternate " sensors" if the real ones fail. We are exploring further use of the Observer for real-time embedded diagnostics based on high speed fuzzy logic chips just becoming available. We have established a Fuzzy Inferencing Test Bed for fuzzy logic applications. It uses a set of development tools which allow applications to be built and tested against simulated systems and then ported directly to a high speed fuzzy logic chip. With the Fuzzy Inferencing Test we investigate very high speed fuzzy logic to: isolate faults using static information and early fault information that evolves rapidly in time validate and smooth readings from redundant sensors and smoothly select alternate control modes in intelligent controllers. This paper reports our experience with fuzzy logic in these kinds of applications.

This paper presents a new fuzzy validity function which is mathematically justified via its relationship to the separation index a well defmed hard clustering validity function. The condition for existence of a unique globally optimal fuzzy c-partition has been found. The performance of this function compares well with other validity measures.

An application of the theory of fuzzy sets to detect and measure convex objects in an image is described. Geometric measurements involving the concept of the perimeter of a fuzzy set are compared to measurements using moment parameters of the membership function. The concept of the perimeter of fuzzy sets offers a way to take geometric measurements from a scene without having to segment it. A method to compute the perimeter of a convex fuzzy set was proposed by Rosenfeld [1]. For the special case of elliptically shaped convex objects an alternative formula is proposed. In this method the fuzzy set is approximated by a crisp set of elliptic shape which has same area and second order moments. The computation of the membership function plays a key role in this theory. We use a fuzzy c-means clustering algorithm to compute the membership function. The method is tested on real images.

We describe the least median of squares (LMedS) robust estimator which identifies the surface corresponding to the absolute majority of the data points. However when all the data points are corrupted by noise LMedS may fail. This is the case in computer vision applications and we have developed a new approach which preserves the robustness of LMedS but avoids its artifacts in the presence of noise.

This paper describes a procedure for segmentation of color face images. A cluster analysis algorithm uses a subsample of the input image color pixels to detect clusters in color space. The clustering program consists of two parts. The first part searches for a hierarchical clustering using the NIHC algorithm. The second part searches the resultant cluster tree for a level clustering having minimum description length (MDL). One of the primary advantages of the MDL paradigm is that it enables writing robust vision algorithms that do not depend on user-specified threshold parameters or other " magic numbers. " This technical note describes an application of minimal length encoding in the analysis of digitized human face images at the NTT Human Interface Laboratories. We use MDL clustering to segment color images of human faces. For color segmentation we search for clusters in color space. Using only a subsample of points from the original face image our clustering program detects color clusters corresponding to the hair skin and background regions in the image. Then a maximum likelyhood classifier assigns the remaining pixels to each class. The clustering program tends to group small facial features such as the nostrils mouth and eyes together but they can be separated from the larger classes through connected components analysis.

Developing techniques for interpreting the structure of volumetric data is useful in many applications. A key initial stage is volumetric segmentation involving the processes of pariilionirig and idenificaion. Here we present a parallel algorithm which through the use of a-partitioning and volume filtering segments volumetric image data such that the greylevel variation within each volume can be described by a regression model. Experimental results demonstrate the effectiveness of this algorithm on several real-world 3D images.

Image segmentation involves labelling pixels according to their membership in image regions. This requires that we understand what a region is. Using supervised pixel classification, we investigate how groups of pixels labelled manually according to perceived image semantics map onto the feature space created by an Artificial Visual System. We investigate multiscale structure of regions and show that pixels form clusters based on their geometric roles in the image intensity function, not by image semantics. A tentative abstract definition of a "region" is proposed based on this behavior.

This paper outlines the framework of an image segmentation system based on the expected presenialion of objects in an image. The paradigm uses models that best characterize those objects that are likely to be present in a scene as captured by a given image formation process. We present the parameters for describing the expected presentations and show how they can be developed into a regionbased image algebra that is a generalized mechanism for reasoning and planning image segmentation and subsequent machine learning tasks. We present results of experiments with Transmission Electron Microscope (TEM) serial sections aerial photographs of urban scenes Mill brain scans and dental radiographs.

A segmentation approach is presented that uses combined information of range data and structured light. Structured light are good for extracting surface patches that cannot be easily obtained from gray scale images. Furthermore surface normal discontinuity can be detected at light stripe flection points. Accidental alignments of light stripes may confuse the boundaries of surfaces that are at different depths. This problem can be solved using range data that are specially good for detecting depth discontinuity.

Image segmentation techniques which employ domain knowledge produce in general much better results than context-free methods but often suffer from lack of portability and excessive development and computation time. We describe a new segmentation technique which employs very general knowledge of domain characteristics to improve the performance of context-free systems while avoiding many of the problems of existing domain-specific methods.

This paper introduces a method for range image segmentation which combines the useful properties of edge and region-based approaches. In the region-based approach the partial derivatives are first estimated at every point on the image by fitting a quadnc model to a small neighborhood of pixels. Pixels are classified into 10 surface types according to the spatial properties in the neighborhood of each pixel. Then pixels of the same surface type are grouped into geometrically coherent regions. Edge detection employs a two stage method which detects both step and roof edges. Finally the edge and region-based segmentations are combined to form the final segmentation and a set of geometrical features is calculated for each region to be used in further analysis.

In this work we propose a segmentation algorithm working on surface normal information. An edge-based coarse segmentation map is computed by the detection of discontinuities in surface orientations. A region-based segmentation is generated by the analysis of surface curvature. A decision making process produces the final segmentation.

This paper describes an original approach to the analysis and prediction of graylevel textures generated as equilibrium states of Gibbs/Markov random fields. This approach is physically motivated by the analogy that exists between the graylevel textures and the miscibility patterns of multiphase flows. The physics of the situation is captured using miscibility matrices that are related to the co-occurrence matrices classically used for texture discrimination. Simulations are provided to motivate and illustrate our approach.

We give an introduction to the phenomenon ofsymmetry and briefly describe its importance to machine vision tasks. An algorithm is then presented which finds all axes of local symmetry builds these axes into the axes of subparts and then uses this new set of axes to approximate image features with symmetrical shapes. Finally a hierarchy of such axes is produced.

A generalized Lloyd algorithm having a tuning parameter for image segmentation is investigated whose algorithm is derived from a minimum weighted squared distortion criterion to select a suitable threshold from histogram information. We consider how to control the parameter of the algorithm by supervised approaches to realize reliable segmentation. Two approaches based on a probabilistic histogram model and a typical image in processing images are treated in detail.

Most previous work on image texture has not studied the effect of noise on texture classification or on textured image segmentation. Classification studies typically use samples of unknown textures that were digitized under the same noise-free conditions as the samples in the training set . In this paper noise-tolerant texture classification is demonstrated. Noise-tolerant texture classification allows the samples of unknown textures to be digitized under different noise and illumination conditions than the samples in the training set. The features extracted from the texture samples are relatively tolerant of noise and illumination gradients resulting in reliable classification. The method is based on edges found in texture samples using a noise-tolerant edge detector similar to the Canny operator [4]. Texture features are average distances separating edges in various orientations [10 11]. Since the edges are reliably extracted from noisy samples texture features based on these edges are noise-tolerant. This method is investigated experimentally and compared with cooccurrence matrix features [6] and cooccurrence matrix features of directionally-smoothed samples. The edge-based features may be used for texture-based image segmentation. This is demonstrated by partitioning an image into areas of uniform texture even when there are illumination and noise gradients over the image.

Texture is an important property useful for image segmentation and the inference of 3-D information in the scene. Many approaches were proposed for analyzing textures. Among them are feature-based approaches and model-based approaches. In a feature-based environment various textural features are extracted from each textured image(or subimage) and are used to classify or discriminate given textures i. e. no explicit consideration of models is taken into account and thus the generation aspect is ignored. In model-based analysis we describe texture in terms of mathematical model which has both analysis and synthesis abilities. In the literature several comparative studies of feature-based methods are found. However few explicit comparative studies of model-based methods have been reported. This paper describes the development of some criteria to compare two model-based texture analysis methods (Time Series model and Markov Random Field model).

One of the major problems in object recognition is finding good candidate regions to start the recognition process. One way of selecting a region in the image is to find one that is distinctive in some feature such as color texture etc. In this paper we present an approach to finding distinctive colored regions in an image. It is based on the observation that although humans can discriminate a large number of colors psychophysically they can partition the color space into only a relatively few distinct qualitative color sensations. The algorithm first builds a color feature map that describes the image as consisting of hierarchical overlapping colored regions. The distinctiveness of these regions is then judged using measures such as size individual color contrast etc.

A method is described for the proe-processing of colour images. The histogram of a colour image is scanned using a Peano scan and the resulting one dimensional histogram is binned into 256 bins in a method similar to histogram equalisation. New co1ours are calculated for these bins and substituted back into the original image. The method produces high quality reproductions of the original and allows further processing to be achieved at an increased rate.

Range images incorporate 3-D surface coordinates of a scene and are well suited for a variety of vision applications. For tasks such as 3-D object recognition a representation of the object(s) present in the image is derived and then matched with stored models to determine the object(s) identity. Surface based representation is the most widely used representation in range image analysis. To generate surface based representation of an object a segmentation of the object into a number of surfaces is needed. In this paper we present an approach for the segmentation of range images into a number of surfaces that in turn can be used to generate surface based representation. The approach integrates both edge detection and region growing techniques to achieve the segmentation. We start by detecting jump edges. Jump edge map is processed and regions surrounded by jump edges are isolated. Next fold edges are detected iteratively using normals and residual. Fold edge map is processed to obtain the final segmented image. Jump and fold edge maps are processed using a Bayesian approach. The apriori knowledge is modeled using Markov Random Field. For jump edges we have used a coupled line and depth process. For fold edges we have combined line residuals and normals to process the fold edge map. The performance of the algorithm on a number of range images is presented.

A neural network approach to segmentation of forward looking infrared and synthetic aperture radar imagery is presented. This approach integrates three stages of processing. First a wavelet transform of the image is performed by projection of the image onto a set of 2-D Gabor functions. This results in a multiple-resolution decomposition of the image into oriented spatial frequency channels. Scond a neural network optimization procedure is used to estimate the wavelet transform coefficients. The third stage involves a segmentation technique that has been shown to work well on textures that human subjects readily segment into regions. Although the approach is still under development preliminary results are promising. The direction of further research efforts are discussed.