The rank-order hit-miss transform (HMT) filter is a significant new advancement in pattern recognition. We detail a new version of our HMT algorithm used for detection, the filter parameters used, and detection (P_D) and false alarm (P_FA) results. In detection, these filters are required to locate all objects in a scene with clutter present. This must be achieved for objects in multiple different classes, with 3-D distortion and contrast differences present. Thus, they represent considerably new image processing filters.

This paper deals with state-of-the-art novel ideas in high level visualization and understanding 3D line drawing objects. A new model-based strategy is presented. It uses only very few learning samples and is more accurate than the conventional linear combination method.

New manufacturing technologies for rapid prototyping and manufacturing have remarkably evolved in recent years. Especially the new laser technology named lasercaving, where three dimensional geometries are directly realizable, gives rise to new types of applications. If it comes to the point where digitized data are needed, for rapid prototyping and manufacturing, we can no more afford to do this `by hand.' Therefore image processing systems with measurement features are needed to support the digitization and inspection processes. Related to the lasercaving method the usefulness of integrated image processing technologies is described. Application examples are given that highlight the power of the integrated system lasercaving and image processing.

Feature extraction and object representation form the basis of a model based object recognition system. This paper presents a new technique for invariant representation of object geometry. A massively parallel line feature extractor is applied to an edge extracted image to detect the object features. A feature tree representation of the object geometry is generated. The feature tree thus becomes a complete representation of the objects in the image. A line based object representation technique has been chosen over the point based technique to circumvent the uncertainty associated with point representation. The geometric invariants of the lines are computed using a combination of line features. An invariant feature tree representing the object geometry is thus obtained. This representation is highly parallel and can be effectively exploited in the recognition phase.

Machine vision technology has got a strong interest in Finnish research organizations, which is resulting in many innovative products to industry. Despite this end users were very skeptical towards machine vision and its robustness for harsh industrial environments. Therefore Technology Development Centre, TEKES, who funds technology related research and development projects in universities and individual companies, decided to start a national technology program, Machine Vision 1992 - 1996. Led by industry the program boosts research in machine vision technology and seeks to put the research results to work in practical industrial applications. The emphasis is in nationally important, demanding applications. The program will create new industry and business for machine vision producers and encourage the process and manufacturing industry to take advantage of this new technology. So far 60 companies and all major universities and research centers are working on our forty different projects. The key themes that we have are process control, robot vision and quality control.

One of the central problems of computer vision is model based object recognition. A catalogue of model objects is described via a set of features which are matched with the images. Image features that are most important to object recognition can be categorized into two types: edges and regions. Edges and surface patches are obtained from intensity and range images via independent algorithms. In this paper, we describe an algorithm that sets up a correspondence between the edge and surface features in a multiobject scene. Each edge is labeled with a list of surfaces to which it belongs and each surface is labeled with a list of edges that are on its boundary. This information can then be passed to higher level routines to group into individual objects for scene analysis. Both synthetic (with Gaussian noise) and real images containing multiple object scenes have been tested. Results appear quite encouraging.

This paper describes the use of intelligent image processing as a machine guarding technology. One or more color, linear array cameras are positioned to view the critical region(s) around a machine tool or other piece of manufacturing equipment. The image data is processed to provide indicators of conditions dangerous to the equipment via color content, shape content, and motion content. The data from these analyses is then sent to a threat evaluator. The purpose of the evaluator is to determine if a potentially machine-damaging condition exists based on the analyses of color, shape, and motion, and on `knowledge' of the specific environment of the machine. The threat evaluator employs fuzzy logic as a means of dealing with uncertainty in the vision data.

We present a model-based 3D object recognition architecture that combines pose estimation derived from range images and hypothesis verification derived from intensity images. The architecture takes advantage of the geometrical nature of range images for generating a number of hypothetical poses of objects. Pose and object models are then used to reconstruct a synthetic view of the scene to be compared to the real intensity image for verification. According to the architecture a system has been implemented and successful experiments have been performed with boxes of different shapes and textures. Recognition with our approach is precise and robust. In particular verification can detect false poses resulting from wrong groupings. In addition, the system provides the interesting features to recognize the true pose of shape-symmetrical objects and also to recognize objects that are ambiguous from their sole shape.

Planetary exploration is a new challenge in the field of space robotics. Most of the topographic information available today about planetary surfaces was gathered either by orbiters or stationary landers. The use of an autonomous rover capable of building detailed maps of unstructured natural terrains would be more relevant, in the long term, for such space exploration missions. This paper presents a natural terrain model to represent and analyze the information gathered by an exploration rover using 3D computer vision. Irregular triangular meshes are suitable for the representation of amorphous natural terrain and could easily adapt to its irregular nature. The triangular terrain model presented captures both the range and the intensity information of the planetary scenes, enabling the modeling of topographic features at different resolution levels and the construction of conceptual topologic maps. In this paper, we first provide a compact representation for the rugged terrain scenes, and then use the resulting topographic maps to assist visual navigation of an autonomous exploration rover. The symbolic topologic maps can be used for the strategic planning of the exploration mission and reasoning about its various alternatives.

We describe procedures that extract line drawings from digitized gray level images, without use of domain knowledge, by modeling preattentive and perceptual organization functions of the human visual system. First, edge points are identified by standard low-level processing, based on the Canny edge operator. Edge points are then linked into single-pixel thick straight- line segments and circular arcs: this operation serves to both filter out isolated and highly irregular segments, and to lump the remaining points into a smaller number of structures for manipulation by later stages of processing. The next stages consist in linking the segments into a set of closed boundaries, which is the system's definition of a line drawing. According to the principles of Gestalt psychology, closure allows us to organize the world by filling in the gaps in a visual stimulation so as to perceive whole objects instead of disjoint parts. To achieve such closure, the system selects particular features or combinations of features by methods akin to those of preattentive processing in humans: features include gaps, pairs of straight or curved parallel lines, L- and T-junctions, pairs of symmetrical lines, and the orientation and length of single lines. These preattentive features are grouped into higher-level structures according to the principles of proximity, similarity, closure, symmetry, and feature conjunction. Achieving closure may require supplying missing segments linking contour concavities. Choices are made between competing structures on the basis of their overall compliance with the principles of closure and symmetry. Results include clean line drawings of curvilinear manufactured objects. The procedures described are part of a system called VITREO (viewpoint-independent 3-D recognition and extraction of objects).

We introduce an approach to the representation of curved or polyhedral 3-D objects and apply this representation to pose estimation. The representation is based on surface patches with uniform curvature properties extracted at multiple scales. These patches are computed using multiple alternative decompositions of the surface based on the signs of the mean and Gaussian curvatures. Initial coarse decompositions are subsequently refined using a curvature compatibility scheme to rectify the effect of noise and quantization errors. The extraction of simple uniform curvature features is limited by the fact that the optimal scale of processing for a single object is very difficult to determine. As a solution we propose the segmentation of range data into patches at multiple scales. A hierarchical ranking of these patches is then used to describe individual objects based on geometric information. These geometric descriptors are ranked according to several criteria expressing their estimated stability and utility. The applicability of the resulting multi-scale description is demonstrated by estimating the pose of a 3-D object. Pose estimation is cast as an optimal matching problem. The geometric pose transformation is computed by matching two representations, which amounts to finding the three-patch correspondence that produces the best global consistency. Examples of the multi- scale representation applied to both real and simulated range data are presented. Effective pose estimation is demonstrated and the algorithm's behavior in the presence of noise is validated.

Theoretical studies of visual form perception have proposed hierarchical representations of three dimensional shape as a basis for achieving fast and reliable recognition of a wide range of objects. In this article we relate an earlier developed representation of object shape to two different types of recognition processes: (1) recognition of familiar objects, and, (2) recognition of the possible uses or affordances of not necessarily known objects in actions. The representation is based on the connectedness and neighborhood relations of object shape. It consists of three hierarchy levels of parts, sub-parts and surface patches, which build topologies of increasing strength. Each level has an associated set of qualitative and quantitative features. We submit that the visual knowledge needed for recognizing a known object is made explicit primarily at the part level and knowledge about affordances at the sub- part level. Recognition of the possible uses of objects is treated as finding the compatibility between action requirements and object affordances. The possibility and necessity of fuzzy sets are used as measures for the compatibility of the individual requirement-affordance pairs and their aggregation to the overall compatibility of a given object and action.

Edge extraction and chaining of the edgels often don't use external information and are automatic processes which rely on some manually adjusted parameters. This paper shows that in many cases external information exits and can be used to improve edge extraction and chaining. It also shows how this can be done and gives a few examples. External information can take many forms: a brief description of the scene from which the image is taken, results of the processing of another image of the same scene, knowledge of the sensor. Instead of using a `blind process' which refers to fixed parameters and which considers every pixel in the same way, the idea of the paper is to exploit the external information during all the phases of the image processing. The paper describes different cases of external information usage and explains how this information can be used to improve edge, chain or segment extraction. Examples are given both in infrared and MMW domains.

Edges are important features in image analysis and edge detection by use of Gaussian filters is widely used in image processing and computer vision. After analyzing the problems of the classical edge detection methods using Gaussian filters, we propose in the present paper a fast Gaussian-filter-based edge detection method, called Hermite integration method, by use of Hermite polynomial theory. One-dimensional Gaussian filtering is at first analyzed and the fast algorithm by use of the input signal samples corresponding to the Hermite polynomial roots is proposed. To use this new algorithm for edge detection in noisy 2-D images, we generalize then this method to Gaussian-filtered derivatives calculation and to multi-dimensional cases, such as 2-D image processing. We show as well that with the proposed method, one can detect the edges with a subpixel precision and calculate the image characteristics for any subpixel positions, which is difficult for classical methods. Our algorithm gives a better algebraic precision and a less important complexity than the classical mask convolution method. Experimental results for artificial data and real images are reported, which confirm the theoretical analysis.

We present a new method for linking edge points in a digital image, and segmenting the resulting edges into simple geometric elements. The initial linking procedure operates on the raw output of conventional edge detection algorithms, and links the pixels into sequences in a manner that guarantees the absence of branches. This linking requires no computationally expensive directional calculations to minimize branching. The resulting contours can, however, be of arbitrary length and complexity. In order to facilitate their later manipulation by higher-level algorithms, these contours are then segmented into straight line segments and circular arcs. The segmentation procedure relies on the overall symmetry of the detected segments, and avoids problems associated with the detection of corners or high curvature points. A contour segment is said to possess the considered overall symmetry property if, within certain tolerances, for any point on the segment, travelling an equal distance along the segment on each side of the point leads to contour points separated by equal straight-line chords from the central point. It appears that within the framework of digitized images, this property can only be satisfied by straight line segments and circular arcs. We describe an algorithm to extract, from arbitrary non-branching contours, segments verifying this symmetry property. After extraction, segments are classified as lines or arcs, and the radii and centers are estimated for the latter.

The human visual system uses two-dimensional (2D) boundary information to recognize objects since the shape of the boundary usually contains the pertinent information about an object. Thus, representing a boundary concisely and consistently is necessary for object recognition. In this paper, we propose a consistent object representation method using mean field annealing (MFA) technique for computer vision applications. Since a curvature function computed on a preprocessed smooth boundary, which is obtained by the MFA approach is consistent, we can consistently detect corner points in this curvature function space. Furthermore, the MFA approach preserves the sharpness of corner points very well. Thus, we can detect corner points easier and better with this method than with other existing methods. Ideal corner points rarely exist for a real boundary. They are often rounded due to the smoothing effect of the preprocessing. In addition, a human recognizes both sharp corner points and slightly rounded segments as corner points. Thus, we use `corner sharpness,' which is qualitatively similar to a human's capability of detecting corner points, to increase the robustness of the proposed algorithm.

FIPS (Flexible Image Processing System) and IQIS (Integrated Quality Inspection System) are two components of a new software package, that enable the user to integrate image processing into the manufacturing environment. FIPS is a user friendly software tool. It consists of different user levels that allow for linkage to CAD and CAM systems as well as the adaptation to sensor and robot environments. IQIS delivers the interfaces for the manufacturing environment such as industrial robots, manufacturing cells, etc. By way of example the different features are described.

We describe a prototype system for interactive image processing using Prolog, implemented by the first author on an Apple Macintosh computer. This system is inspired by Prolog+, but differs from it in two particularly important respects. The first is that whereas Prolog+ assumes the availability of dedicated image processing hardware, with which the Prolog system communicates, our present system implements image processing functions in software using the C programming language. The second difference is that although our present system supports Prolog+ commands, these are implemented in terms of lower-level Prolog predicates which provide a more flexible approach to image manipulation. We discuss the impact of the Apple Macintosh operating system upon the implementation of the image-processing functions, and the interface between these functions and the Prolog system. We also explain how the Prolog+ commands have been implemented. The system described in this paper is a fairly early prototype, and we outline how we intend to develop the system, a task which is expedited by the extensible architecture we have implemented.

Deformable models using energy minimization have proven to be useful in computer vision for segmenting complex objects based on various measures of image contrast. In this paper, we incorporate prior shape knowledge to aid boundary finding of 2D objects in an image in order to overcome problems associated with noise, missing data, and the overlap of spurious regions. The prior shape knowledge is encoded as an atlas of contours of default shapes of known objects. The atlas contributes a term in an energy function driving the segmenting contour to seek a balance between image forces and conformation to the atlas shape. The atlas itself is allowed to undergo a cost free affine transformation. An alternating algorithm is proposed to minimize the energy function and hence achieve the segmentation. First, the segmenting contour deforms slightly according to image forces, such as high gradients, as well as the atlas guidance. Then the atlas is itself updated according to the current estimate of the object boundary by deforming through an affine transform to optimally match the boundary. In this way, the atlas provides strong guidance in some regions that would otherwise be hard to segment. Some promising results on synthetic and real images are shown.

A robust segmentation scheme is one of the primary requirements for three-dimensional object recognition. The task of partitioning a given image into homogeneous regions has been the centerpiece of investigations of several major researchers all of these years. In this paper we propose a simplistic range image segmentation scheme for un-occluded cluttered three- dimensional objects in a scene. An assumption relating to the three-dimensional objects being quadric in nature, would further enable us to demonstrate an object recognition scheme which has been proposed in an earlier publication. The proposed method involves the extraction of jump edges which we refer to as global edges. After having median filtered the resultant image, a thinning algorithm is implemented which is subsequently followed by the blob determination algorithm thereby completing the segmentation process. Experiments have been conducted on range images of scenes consisting of several quadric surfaces with promising results. Application of the object recognition scheme subsequently successfully classified most of the objects in the scenes as either spherical, cylindrical, or planar in nature.

Analysis of range data is important for performing scene interpretation in robotic environments. The first step applied to range images for feature extraction is segmentation, which reveals the inner borders between three-dimensional scene elements. Segmentation of range images can be performed by detecting edges and interpreting them as boundaries of the different surfaces. A segmentation method for range images is proposed using the a trous algorithm implementation of the discrete wavelet transform. This algorithm applies oriented band pass filters at multiple scales for derivative estimations. The wavelet transform (WT) is used as a smoothing multiscale differentiation operator. A model for range image features is developed based on the differential properties of step and roof edges in range images. We show how specific families of wavelets are used for investigating the local and scale properties of the differentials of two-dimensional signals, with emphasis on the higher derivative singularities and zero-crossings usually found in range data. These features are detected, then combined into a binary edge map. Following the modulus maxima of the wavelet transform in the direction of the gradient vector obviates the need for thresholding. The resulting binary edge map provides a basis for complete segmentation. The final result is a segmented image with labeled regions indicating different surface patches. This segmented image can be used as input to a higher-level recognition or for a three-dimensional reconstruction algorithm. The technique is applied to synthetic and real range images with different features and is shown to yield consistent and reliable results.

Gabor transform has recently been exploited to do texture analysis, including texture edge detection, texture segmentation/discrimination, and texture synthesis. For most of the applications using Gabor transform, people convolve the given texture image with a set of Gabor filters with some user specified parameters. Although the mathematical formulation of applications involve the Fourier transform, few have investigated mathematical properties of the relationship between Gabor filters and their Fourier transform. This paper mainly studies mathematical properties of real Gabor filters and their corresponding Fourier transform. The goal is to select a set of `interesting' Gabor filters, or say, a set of parameters for Gabor filters to do texture analysis. We demonstrate, by means of 3-D graphical displays, that a Gabor filter or its corresponding Fourier transform may have a single peak or double peaks according to different parameters. Experiments for texture discrimination are given to demonstrate the applications of Gabor transform.

The formulation of object hypotheses for recognition requires the localization of the most meaningful groups of lines in the image, that correspond to the elementary structures used to describe and represent objects or some parts of them. This paper describes an approach where the most salient line segments are used to suggest main structures. A measure of saliency is proposed on the basis of length and luminance contrast of line segments. In order to enhance the performance of the algorithm, only a reduced number of line segments are taken into account to formulate rough structures. These significant lines are ordered according to their saliency and the contour map is analyzed in a coarse to fine sequence. Then a grouping strategy based on perceptual organization criteria extracts closed polygons, paying special attention to quadrangles and triangles, C-shaped and L-shaped structures. The saliency measure allows the grouping process to focus on the bigger and/or more evident structures, giving priority to a coarse aggregation. Besides a fine aggregation is performed at the same time to reinforce and refine each coarse aggregation step. This cooperation allows a sophisticated usage of spatial thresholds, that reduces the direct impact of specific threshold values making grouping process less sensitive to the scale of the objects present in the image.

We are suggesting a new approach to NMR image segmentation for multiple scleroses volume quantification. The choice of a segmentation technique is conditioned by what is known in NMR acquisition artifacts. This research has been conducted in collaboration with the Lille Neuroradiology Laboratory leading to the implementation of a disease development index.

Shape from shading has been extensively studied in the field of computer vision. Most research in this area assumes a non-attenuating medium, such as air. Generally speaking, the developed methods cannot be directly used for applications in undersea environments where the absorption and scattering of the medium may not be negligible. In this paper, we present an image irradiance model for a Lambertian surface illuminated by a point source in an attenuating medium such as seawater. This model has been used to recover the orientation of uniform planar patches. Our current experimental data show that the developed model also provides a quite accurate approximation of the image brightness of curved surfaces. In this paper, we discuss several ambiguities in the solutions based on this model and compare them with solutions in land-based systems. Our discussion mainly involves case studies of ruled surface which can be reduced to 1-D cases in the computation. The results presented in this paper are useful for image understanding and/or object recognition in underwater applications.

A self-organizing neural network is developed for recognition of 3-D objects from sequences of their 2-D views. Called VIEWNET because it uses view information encoded with networks, the model processes 2-D views of 3-D objects using the CORT-X 2 filter, which discounts the illuminant, regularizes and completes figural boundaries, and removes noise from the images. A log-polar transform is taken with respect to the centroid of the resulting figure and then re-centered to achieve 2-D scale and rotation invariance. The invariant images are coarse coded to further reduce noise, reduce foreshortening effects, and increase generalization. These compressed codes are input into a supervised learning system based on the Fuzzy ARTMAP algorithm which learns 2-D view categories. Evidence from sequences of 2-D view categories is stored in a working memory. Voting based on the unordered set of stored categories determines object recognition. Recognition is studied with noisy and clean images using slow and fast learning. VIEWNET is demonstrated on an MIT Lincoln Laboratory database of 2-D views of aircraft with and without additive noise. A recognition rate of up to 90% is achieved with one 2-D view category and of up to 98.5% correct with three 2-D view categories.

A new classifier neural network is described for distortion-invariant multi-class pattern recognition. Its input training data in different classes are described by a feature space. As a distortion parameter (such as aspect view) of a training set object is varied, an ordered training set is produced. This ordered training set describes the object as a trajectory in feature space, with different points along the trajectory corresponding to different aspect views. Different object classes are described by different trajectories. Classification involves calculation of the distance from an input feature space point to the nearest trajectory (this denotes the object class) and the position of the nearest point along that trajectory (this denotes the pose of the object). Comparison to other neural networks and other classifiers show that this feature space trajectory neural network yields better classification performance and can reject non-object data. The FST classifier performs well with different numbers of training images and hidden layer neurons and also generalizes better than other classifiers.

Analysis of catastrophic forgetting by distributed codes leads to the unexpected conclusion that the standard synaptic transmission rule may not be optimal in certain neural networks. The distributed outstar generalizes the outstar network for spatial pattern learning, replacing the outstar source node with a source field, of arbitrarily many nodes, where the activity pattern may be arbitrarily distributed or compressed. Distributed outstar learning proceeds according to a principle of atrophy due to disuse whereby a path weight decreases in joint proportion to the transmitted path signal and the degree of disuse of the target node. During learning, the total signal to target node converges toward that node's activity level. Weight changes at a node are apportioned according to the distributed pattern of converging signals. Three types of synaptic transmission, the standard product rule, a fuzzy capacity rule, and an adaptive threshold rule, are examined for this system. Only the threshold rule solves the catastrophic forgetting problem of fast learning. Analysis of spatial distributed coding hereby leads to the conjecture that the unit of long-term memory is a spatial pattern learning system may be a subtractive threshold, rather than a multiplicative weight.

During recent years, significant advances have been made in two distinct technological areas: fuzzy logic and computational neural networks. The theory of fuzzy logic provides a mathematical framework to capture the uncertainties associated with human cognitive processes, such as thinking and reasoning. It also provides a mathematical morphology to emulate certain perceptual and linguistic attributes associated with human cognition. On the other hand, the computational neural network paradigms have evolved in the process of understanding the incredible learning and adaptive features of neuronal mechanisms inherent in certain biological species. Computational neural networks replicate, on a small scale, some of the computational operations observed in biological learning and adaptation. The integration of these two fields, fuzzy logic and neural networks, have given birth to an emerging technological field -- fuzzy neural networks. Fuzzy neural networks, have the potential to capture the benefits of these two fascinating fields, fuzzy logic and neural networks, into a single framework. The intent of this tutorial paper is to describe the basic notions of biological and computational neuronal morphologies, and to describe the principles and architectures of fuzzy neural networks. Towards this goal, we develop a fuzzy neural architecture based upon the notion of T-norm and T-conorm connectives. An error-based learning scheme is described for this neural structure.

A RAM-based neural network applicable for object detection in machine vision is considered. It is shown that it is easy to perform a crossvalidation test for the training set using this network type. This is relevant for measuring the network generalization capability (robustness). An information measure combining the concept of crossvalidation and Shannon information is proposed. We describe how this measure can be used to select the input connections of the network. The task of recognizing handwritten digits is used to demonstrate the capability of the selection strategy.

In this paper, we propose a Gaussian neuron model for feedforward type of neural networks and a method to adapt the above network for any input, not necessarily in the range [0,1]. An error function based on the class label and a priori probability is defined and gradient descent procedure, with backpropagating error, is used for finding the optimal set of parameters of this network. Different approaches are proposed for increasing the rate of convergence of this network. Experimental results are given for continuous data from speech waveform and XOR type of data.

The ability to rapidly detect moving objects while dynamically exploring a work environment is an essential characteristic of any active vision system. However, many of the proposed computer vision paradigms are unable to efficiently deal with the complexities of real world situations because they employ algorithms that attempt to accurately reconstruct structure- from-motion. An alternative view is to employ algorithms that only compute the minimal amount of information necessary to solve the task at hand. One method of qualitatively detecting independently moving objects by a moving camera (or observer) is based on the notion that the projected velocity of any point on a spherical image is constrained to lie on a one-dimensional locus in a local 2-D velocity space. The velocities along this locus, called a constraint ray, correspond to the rotational and translational motion of the observer. If the observer motion is known a priori, then any object moving independently through the rigid 3- D environment will exhibit a projected velocity that does not fall on this locus. As a result, the independently moving object can be detected using a clustering algorithm. In this paper, a hybrid neural network architecture is proposed for discriminating between flow velocities that are caused by camera movement and by object motion. The computing architecture is essentially a two stage process. In the first stage, a self-organizing neural network is used to learn the constraint parameters associated with typical observer movements by moving the camera apparatus through a stationary environment. Once the observer movements have been adequately learned by the self-organizing neural network, the corresponding synaptic weight values are used to program a modified radial basis function (RBF) network. During the second stage, the RBF network architecture acts as a constraint region classifier by employing clustering strategies to label incomplete motion field information (i.e. the velocity component that is parallel to the spatial gradient).

In this paper we present a hybrid neural method that uses a neural network to generate initial search points for a discrete heuristic. We demonstrate the method for the subset-sum problem (SSP). The method hinges on using the continuous valued activations of the neural system to select a corner of the n-cube that can be used to initialize a discrete search. This can be done at each neural iteration, resulting in many discrete searches over the coarse of a single neural run. Without the discrete heuristic, the selected corners can be interpreted as instantaneous neural solutions and the best-so-far tabulated as the neural system runs. This allows the neural system to be terminated without losing the full effort of the run, and should the network be run until convergence, the best-so-far result is at least as good as the convergent corner, and usually better. With the discrete heuristic, a search is launched from the instantaneous neural solutions, greatly improving the overall results (again keeping the best-so-far). The results are presented for an n equals 25 SSP, with comparisons to simulated annealing and genetic approaches.

A goal of computer vision is the construction of scene descriptions based on information extracted from one or more 2D images. A reconstruction strategy based on a four-level representational framework is presented. We are interested in the second representational level, the Primal Sketch. It makes explicit important information about the two-dimensional image, primarily the intensity changes and their geometrical distribution and organization. The intensity changes corresponding to physical features of the observed scene appear at several spatial scales, in contrast to spurious edges, and image analysis performed at multiple resolutions is therefore more robust. We propose a compact pyramidal neural network implementation of the multiresolution representation of the input images. Features of the scene are detected at each resolution level and feedback interaction is built between pyramid levels in order to reinforce edges which correspond to physical features of the observed scene. A vigilance neuron determines the importance granted to each spatial resolution in the feature extraction process.

In the modern sawmill industry automatic grading of the products is one of the key issues in increasing the production quality. The surface defects that determine the grading are identified according to the physiological origin of the defect, such as dry, encased or decayed knot. Variations within the classes are large since the knots can have different shapes, sizes and color, and each class has different discriminating features. Classification of the defects using pattern recognition techniques has turned out to be rather difficult, since it is difficult to determine the suitable features that would correlate with the physiological defect types. In this paper we describe a wood defect classification system that is based on self-organizing feature construction and neural network classification. Due to the automatic, unsupervised learning of the features, the system is easily adaptable to different tasks, such as inspection of lumber or veneer, with different tree species and different cutting processes. Performance of the classification system was evaluated with a set of over 400 samples from spruce boards. The knot recognition rate was about 85% with only gray level images, giving about 90% accuracy for the final board grading. Compared to 75 - 80% accuracy that can be maintained by a human inspector, the result can be considered good.

A robot based intelligent system is proposed and used to improve the accuracy and throughput rate of a dynamic checkweigher. Classical filtering techniques as well as some other signal processing techniques provide certain improvements in the accuracy and the effect of high frequency noise in a conventional checkweigher. A large amount of inaccuracy from system low frequency components still remains. The developed system includes a fuzzy controller for the weighing cell which is an essential pat of many static and dynamic systems used for weighing. The system also includes a robot arm for package handling through the weighing process to reduce the effect of the low frequency noise (0 - 10 HZ) associated with the conveyor belt systems in the conventional checkweighers. A motion planning for the robot arm is investigated to satisfy the safety requirements for the packages and robot arm and to enhance the throughput rate of the overall system. The use of such intelligent systems for weighing and transport overcome the nonlinearity problems associated with the system and reduces greatly the noise effect in conventional checkweighers. The experimental results are introduced an analyzed to investigate the efficiency of the developed system.

One of the keys to the successful relational modeling of fuzzy systems is the proper design of fuzzy reference sets. This has been discussed throughout the literature. In the frame of modeling a stochastic system, we analyze the problem numerically. First, we briefly describe the relational model and present the performance of the modeling in the most trivial case: the reference sets are triangle shaped. Next, we present a known fuzzy reference set generator algorithm (FRSGA) which is based on the fuzzy c-means (Fc-M) clustering algorithm. In the second section of this chapter we improve the previous FRSGA by adding a constraint to the Fc-M algorithm (modified Fc-M or MFc-M): two cluster centers are forced to coincide with the domain limits. This is needed to obtain properly shaped extreme linguistic reference values. We apply this algorithm to uniformly discretized domains of the variables involved. The fuzziness of the reference sets produced by both Fc-M and MFc-M is determined by a parameter, which in our experiments is modified iteratively. Each time, a new model is created and its performance analyzed. For certain algorithm parameter values both of these two algorithms have shortcomings. To eliminate the drawbacks of these two approaches, we develop a completely new generator algorithm for reference sets which we call Polyline. This algorithm and its performance are described in the last section. In all three cases, the modeling is performed for a variety of operators used in the inference engine and two defuzzification methods. Therefore our results depend neither on the system model order nor the experimental setup.

As automatic surface inspection systems evolve for metal surfaces, factors affecting defect detection need to be identified and addressed. At Rautaruukki New Technology a system based on a special illumination unit and a CCD-line scan camera has been developed. The Smartvis system uses a novel image compression technique and an adaptive defect detection method in digital image processing. It also features a high performance defect classification method including knowledge extraction techniques and fuzzy set theory based algorithms. The improvement of man-machine interactivity has been one of the major goals in the development. In this paper the factors affecting the decisions made in system design are discussed. Experiences from system applications for steel strip process lines are presented.

During the last few years there has been a large and energetic upswing in research efforts aimed at synthesizing fuzzy logic with neural networks. This combination of neural networks and fuzzy logic seems natural because the two approaches generally attack the design of `intelligent' system from quite different angles. Neural networks provide algorithms for learning, classification, and optimization whereas fuzzy logic often deals with issues such as reasoning in a high (semantic or linguistic) level. Consequently the two technologies complement each other. In this paper, we combine neural networks with fuzzy logic techniques. We propose an artificial neural network (ANN) model for a fuzzy logic decision system. The model consists of six layers. The first three layers map the input variables to fuzzy set membership functions. The last three layers implement the decision rules. The model learns the decision rules using a supervised gradient descent procedure. As an illustration we considered two examples. The first example deals with pixel classification in multispectral satellite images. In our second example we used the fuzzy decision system to analyze data from magnetic resonance imaging (MRI) scans for tissue classification.

In view of the success of neural network applications in inverted pendulum control, speech recognition, and other problem solving, we believe that one could inject the noise removing concepts and learning spirits into the algorithm in constructing the neural networks and apply it to the various tasks such as compliant coordinated motion using multiple robots. Based on the fuzzy logic, a fuzzy logical control system is a logical system which is much closer to human thinking than any other logical systems. During recent years, fuzzy logic control has emerged as a fruitful area in applications, especially the applications lacking quantitative data regarding the input-output relations. Whereas, the connectionist model injects the learning ability to the fuzzy logic system. This model, proposed by Lin and Lee, is a connected neural network that embedded the fuzzy rules in the architecture. Since this model is general enough and we expect the embedded fuzzy concepts can solve the problems caused by the defective training data, it is chosen as our base structure. Appropriate modifications have been made to this model to reflect the real situations encountered in the robot applications. Our goal is to control two different types of robots for coordinated motion using sensory feedback information.

The primary task of machine vision is to utilize a variety of techniques to segment a digital image into meaningful regions, extract the corresponding edges, compute the various features (e.g., area, centroids) and primitives (e.g. lines, corners, curves) that exist in the image, and finally develop some decision rules or grammar structures for interpreting the image content. In conventional vision systems, the operations performed involve making crisp (yes or no) decisions about the regions, features, primitives, regional relationships and overall scene interpretation. However, various degrees of uncertainty exist at each and every stage of the vision system process because these decisions are often based on data that is inexact or ambiguous in nature. Much of the incertitude in the image information can be interpreted in terms of either grayness ambiguity (deciding on the intensity of a pixel) or spatial ambiguity (deciding on the shape and geometry of the regions within the image). Fuzzy morphology is a mathematical tool developed to deal with imprecise or ambiguous information that arises during a subjective evaluation process such as scene interpretation. This mathematical approach transforms a gray scale image into a two-dimensional array of fuzzy singletons called a fuzzy image. The value of each fuzzy singleton reflects the degree to which the pixel possesses some property such as brightness, edgeness, redness, or surface uniformity (i.e., texture). A variety of morphological operations can be performed on the singletons in order to modify the ambiguity associated with the desired property. For the efficient shape representation of objects in a scene, a thinning algorithm for fuzzy images is proposed in this paper. Once the object shape has been thinned to a skeleton-like representation, curve descriptors can be used to transform the generalized shape into a coded form. In essence, this thinning algorithm is used to reduce, or compress, the structural shape information of a vaguely defined object into simplified features for a rule-based description of the object shape.

In this work, we use artificial neural networks to study the problem of reconstructing visual images from their local features. An artificial neural network system with explicit local-feature extraction characteristics was devised to reconstruct visual images. The network studied was a multi-layer feed-forward network, it has a number of special neurons which are designed to resemble the complex and simple cells found in the biological visual systems. The neurons resembling the complex cells extract the lower frequency components of the image and the neurons resembling the simple cells extract the higher frequency components and edge information of the image. The output of these special neurons is forwarded to the higher layers of the network and the network learns to reconstruct the input image from these visually important local features. Experimental results show that excellent quality visual images can be reconstructed from only a few local features. We also discuss the potential applications of such a system to image data compression.

Imprecision can arise in fuzzy relational modeling as a result of fuzzification, inference and defuzzification. These three sources of imprecision are difficult to separate. We have determined through numerical studies that an important source of imprecision is the defuzzification stage. This imprecision adversely affects the quality of the model output. The most widely used defuzzification algorithm is known by the name of `center of area' (COA) or `center of gravity' (COG). In this paper, we show that this algorithm not only maps the near limit values of the variables improperly but also introduces errors for middle domain values of the same variables. Furthermore, the behavior of this algorithm is a function of the shape of the reference sets. We compare the COA method to the weighted average of cluster centers (WACC) procedure in which the transformation is carried out based on the values of the cluster centers belonging to each of the reference membership functions instead of using the functions themselves. We show that this procedure is more effective and computationally much faster than the COA. The method is tested for a family of reference sets satisfying certain constraints, that is, for any support value the sum of reference membership function values equals one and the peak values of the two marginal membership functions project to the boundaries of the universe of discourse. For all the member sets of this family of reference sets the defuzzification errors do not get bigger as the linguistic variables tend to their extreme values. In addition, the more reference sets that are defined for a certain linguistic variable, the less the average defuzzification error becomes. In case of triangle shaped reference sets there is no defuzzification error at all. Finally, an alternative solution is provided that improves the performance of the COA method.

This paper presents a new technique for free form surface reconstruction using functional link neural networks. A comparative study is performed between classical methods used for interpolation and the proposed ones based on neural networks. We show that the neural networks approach provides good approximation for smoothly varying surfaces, and provides better results than classical techniques when used for free form surface reconstruction.

The design and technical capabilities of a true RGB 3 CCD chip color line scan camera are presented within this paper. The camera was developed for accurate color monitoring and analysis in industrial applications. A black & white line scan camera has been designed and built utilizing the same modular architecture of the color line scan camera. Color separation is made possible with a tri-chromatic RGB beam splitter. Three CCD linear arrays are precisely mounted to the output surfaces of the prism and the outputs of each CCD are exactly matched pixel by pixel. The beam splitter prism can be tailored to separate other spectral components than the standard RGB. A typical CCD can detect between 200 and 100 nm. Either two or three spectral regions can be separated using a beam splitter prism. The camera is totally digital and has a 16-bit parallel computer interface to communicate with a signal processing board. Because of the open architecture of the camera it's possible for the customer to design a board with some special functions handling the preprocessing of the data (for example RGB - HSI conversion). The camera can also be equipped with a high speed CPU-board with enough of local memory to do some image processing inside the camera before sending the data forward. The camera has been used in real industrial applications and has proven that its high resolution and high dynamic range can be used to measure minute color differences, enabling the separation or grading of objects such as minerals, food or other materials that could not otherwise be measured with a black and white camera.

Color image processing in machine vision systems has not gained general acceptance. Most machine vision systems use images that are shades of gray. The Laser Automated Decoating System (LADS) required a vision system which could discriminate between substrates of various colors and textures and paints ranging from semi-gloss grays to high gloss red, white and blue (Air Force Thunderbirds). The changing lighting levels produced by the pulsed CO₂ laser mandated a vision system that did not require a constant color temperature lighting for reliable image analysis.

The use of machine vision technology is being investigated at VTT for improving the color quality and productivity of web offset printing. The visual inspection of color quality is performed by a traversing color CCD camera which images the moving web under a stroboscopic light. The measuring locations and goal values for the color register, the ink density, and the gray balance are automatically determined from the postscript description of the digital page. A set of criteria is used to find the best-suited spots for the measurements. In addition to providing data for the on-line control, the page analysis estimates the zone wise ink consumption of the printing plates as a basis for presetting the ink feed. Target colorimetric CIE-values for gray balance and critical colors are determined from the image originals. The on-line measurement results and their deviations from the target values are displayed in an integrated manner. The paper gives test results of computation times, register error measuring with and without test targets and the color measuring capabilities of the system. The results show that machine vision can be used in the on-line inspection of the color print quality.

In this paper a solution is presented which guarantees we avoid the connectivity paradoxes related to the Jordan Curve Theorem for all multicolor images. Only one connectedness relation is used for the entire digital image. We use only 4-connectedness (which is equivalent to 8-connectedness) for every component of every color. The idea is not to allow a certain `critical configuration' which can be detected locally to occur in digital pictures; such pictures are called `well-composed.' Well-composed images have very nice topological properties. For example, the Jordan Curve Theorem holds and the Euler characteristic is locally computable. This implies that properties of algorithms used in computer vision can be stated and proved in a clear way, and that the algorithms themselves become simpler and faster.

Using color for visual recognition outdoors has proven to be a difficult problem, chiefly due to varying illumination. Attempts to classify pixels or image patches in outdoor scenes often fail, partly because of the paucity of the data, but partly because color shifts due to changes in illumination are not well modeled as random noise. Approaches which attempt to recover the `true color' of objects by calculating the color of the incident light (i.e. color-constancy approaches) appear to work in constrained environments, but are not yet applicable to outdoor scenes. We present a technique that uses training images of an object under daylight to learn the shift in color of an object. Our method uses multivariate decision trees for piecewise linear approximation of the region corresponding to the object's appearance in color space. We then classify pixels in outdoor scenes depending on whether they fall within this region, and group clusters of target pixels in to regions of interest (ROIs) for a model-based RSTA system. The techniques presented are demonstrated on a challenging task: recognizing camouflaged vehicles in outdoor military scenes.

This paper describes the relational graph description of natural color scenes and the 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 and regions are then represented using an adjacency relation graph. Scene model is characterized by means of 3-D to 2-D 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.

The question of why the human eye has two axes, a photopic visual axis and an eye axis, is just as justified as the one of why the fovea is not on the eye axis, but instead is on the visual axis. An optical engineer would have omitted the second axis and placed the fovea on the eye axis. The answer to the question of why the design of the real eye differs from the logic of the engineer is found in its prenatal development. The biaxial design was the only possible consequence of the decision to invert the retinal layers. Accordingly, this is of considerable importance. It in turn forms the basis of the interpretation of the retinal nuclear layers as a cellular 3D phase grating, and can provide a diffraction-optical interpretation of adaptive effects (Purkinje shift), aperture phenomena (Stiles-Crawford effects I and II) in photopic vision, and visual acuity data in photopic and scotopic vision.

The notion that color effects in human vision can be explained as diffraction of light by the 3D grating of retina cells was first proposed by N. Lauinger. To study this new diffraction theory of human vision, we solve the wave equation for light diffraction by a 3D-grating layer with rectangular cells, using the method of 4D Fourier spectra. In the case of weak interaction, we derive analytical expressions for the amplitude and intensity of the diffracted light field for the incident plane wave light. The bandwidth of the diffracted light intensity curves is defined by the width of the grating layer, the size of the grating cells, and the grating period. We show that the geometry of the diffracted light is reciprocal with respect to the geometry of the 3D grating. We compute the wavelength dependence of the diffracted light intensity for incident collimated white light for various geometries of the grating layer and the incident light. Within the visible spectrum range 0.4 - 0.7 micrometers , we obtain three main diffracted light intensity curves for the maxima corresponding to red, green and blue colors, which resemble the fundamental sensitivity curves. The behavior of these curves for non-zero incident angle agrees with the Stiles-Crawford effects.

In contrast to differential methods applied to monochromatic images, color edge detection, or multispectral differential methods, necessitates the development of edge operators which can be utilized to determine the presence of a color edge in a number of color spaces. Too often, flat-spaces are used even when the underlying color space is itself non-flat. This is clearly the case with the Hue-Saturation-Intensity (HSI) color space, which describes a complex manifold. Here we concentrate on and explore differential models of the HSI imagery from a differential-geometric viewpoint.

In photopic vision, two physical variables (luminance and wavelength) are transformed into three psychological variables (brightness, hue, and saturation). Following on from 3D grating optical explanations of aperture effects (Stiles-Crawford effects SCE I and II), all three variables can be explained via a single 3D chip effect. The 3D grating optical calculations are carried out using the classical von Laue equation and demonstrated using the example of two experimentally confirmed observations in human vision: saturation effects for monochromatic test lights between 485 and 510 nm in the SCE II and the fact that many test lights reverse their hue shift in the SCE II when changing from moderate to high luminances compared with that on changing from low to medium luminances. At the same time, information is obtained on the transition from the trichromatic color system in the retina to the opponent color system.

Since the 1970s land remote sensing images have been widely used in environment protection, territory management, military, and so on. Moreover, the remote sensing images provide the whole body of information of some regions like mountains, rivers, towns, villages, oceans and other land objectives. How do we extract the significant parts we recovered from the remote sensing images? This paper demonstrates use of the single value feature extraction method, which uses the single value decomposition based on mesh samples to obtain the eigenvalues of the image matrixes. The SPOT satellite remote sensing images are used as models. The size of the meshes are defined by 100 X 100 pixels. Using this method, we experimented on four typical regions and acquired optimal vectors and succeeded in the recognition and classification of the regions. Also, this method has a high speed in computation. The above listed reasons proved that the single value decomposition method is efficient in the classification and the recognition of the typical regions of remote sensing images.

Following