In this paper, two eigenspace transformations are examined for feature extraction and dimensionality reduction in an automatic target detector. The transformations considered in this research are principal component analysis and the eigenspace separation transform (EST). These transformations differ in their capabilities to enhance the class separability and to compact the information for a given training set. The transformed data, obtained by projection of the normalized input images onto a chosen set of eigentargets, are fed to a multilayer perceptron (MLP) that decides whether a given input image is a target or clutter. In order to search for the optimal performance, we use different sets of eigentargets and construct the matching MLPs. Although the number of hidden layers is fixed, the numbers of inputs and weights of these MLPs are proportional to the number of eigentargets selected. These MLPs are trained with a modified Qprop algorithm that maximizes the target-clutter class separation at a predefined false-alarm rate. Experimental results are presented on a huge and realistic data set of forward-looking IR imagery.

We address the problems of recognizing 10 types of vehicles in imagery formed from synthetic aperture radar (SAR). SAR provides all-weather, day, or night imagery of the battlefield. To aid in the analysis of the copious amounts of imagery available today, automatic target recognition (ATR) algorithms, which are either template-based or model- based, are needed. We enhanced template-based algorithms by using an artificial neural network (ANN) to increase the discriminating characteristics of 10 initial sets of templates. The ANN is a modified learning vector quantization (LVQ) algorithm, previously shown effective with forward-looking IR (FLIR) imagery. For comparison, we ran the experiments with LVQ using three different sized temporal sets. These template sets captured the target signature variations over 60 degrees, 40 degrees, and 20 degrees. We allowed LVQ to modify the templates, as necessary, using the training imager from all 10 targets. The resulting templates represent the 10 target types with greater separability in feature space. Using sequestered test imagery, we compared the pre- and post-LVQ template sets in terms of their ability to discriminate the 10 target types. All training and test imagery is publicly available from the Moving and Stationary Target Acquisition and Recognition program sponsored by the Defense Advanced Research Projects Agency.

Determining exactly which neural network architecture, with which parameters, will provide the best solution to a classification task is often based upon the intuitions and experience of the implementers of neural network solutions. The research presented in this paper is centered on the development of automated methods for the selection of appropriate networks, as applied to character recognition. The Network Generating Attribute Grammar Encoding system is a compact and general method for the specification of commonly accepted network architectures that can be easily expanded to include novel architectures, or that can be easily restricted to a small subset of some known architecture. Within this system, the context-free component of the attribute grammar specifies a class of basic architectures by using the non-terminals to represent network, layers and component structures. The inherited and synthesized attributes indicate the connections necessary to develop a functioning network from any parse tree that is generated from the grammar. The attribute grammar encoding is particularly conducive to the use of genetic algorithms as a strategy for searching the space of possible networks. The resultant parse trees are used as the genetic code, permitting a variety of different genetic manipulations. We apply this approach in the generation of backpropagation networks for recognition of characters from a set consisting of 20,000 examples of 26 letters.

The prognosis of malignant melanoma strongly relies on tumor early detection. Unfortunately, differentiating early melanomas from other less dangerous pigmented lesions is a difficult task since they have near physical characteristics. Dermatoscopy is a new non-invasive technique, which, by oil immersion, makes subsurface structures of skin accessible to in vivo examination. Our objective is to develop a computer diagnosis system applied to dermatoscopic images of skin tumors. Most of the signs for the visual diagnosis of melanoma only require the examination of part of the tumors. Our approach consists in classifying windows taken from images of skin tumors by a two-stage procedure. First, a Generalized-Hebbian-Algorithm- based network operates a Principal Component-like Analysis of windows. Sets of primitive windows fitted to various contexts allow both contextual coding and compression of windows. The second stage involves a classical feedforward network, which performs the classification of windows on the basis of the contribution of each primitive window to the reconstruction of windows under consideration. It was shown that classification was properly achieved when 20 primitive windows at least were considered. Application to the classification of skin tumors is in progress and preliminary results dealing with the characterization of borders of lesions are presented.

Neural networks are highly parallel architectures, which have been used successfully in pattern matching, clustering, and image coding applications. In this paper, we review neural network based techniques that have been used in image coding applications. The neural networks covered in this paper include multilayer perceptron (MLP), competitive neural network (CNN), frequency sensitive competitive neural network (FS-CNN), and self-organizing feature map network (SOFM). All of the above mentioned neural networks except MLP are trained using competitive learning and used for designing the vector quantizer codebook. The major problem with the competitive learning is that some of the neurons may get a little or no chance at all to win the competition. This may lead to a codebook containing several untrained codevectors or the codevectors that have not been trained enough. There are several possible ways to solve this problem, FS-CNN and SOFM offer solution to under-utilization of neurons. We present design algorithms for above mentioned neural networks as well as evaluate and compare their performance on several standard monochrome images.

This work presents a novel polynomial neural network approach to 1D differential pulse code modulation (DPCM) design for image compression. This provides an alternative to current tradition and neural networks techniques, by allowing the incremental construction of higher-order polynomials of different orders. The proposed predictor utilizes Ridge Polynomial Neural Networks (RPNs), which allow the use of linear and non-linear terms, and avoid the problem of the combinatorial explosion of the higher-order terms. In RPNs, there is no requirement to select the number of hidden units or the order of the network. Extensive computer simulations have demonstrated that the resulting encoders work very well. At a transmission rate of 1 bit/pixel, the 1D RPN system provides on average a 13 dB improvement in SNR over the standard linear DPCM and a 9 dB improvement when compared to HONNs. A further result of the research was that third-order RPNs can provide very good predictions in a variety of images.

Time-sequential imagery can be acquired by film-based motion camera or electronic video cameras. In this case, there are several factors related to imaging sensor limitations that contribute to the graininess of resulting images. Further, in the case of image sequence compression, random noise increases the entropy of the image sequence and therefore hinders effective compression. Thus, filtering of time- sequential imagery for noise suppression is often a desirable preprocessing step. Some of video image filtering methods use the information about motion in video for reduction of noise. The most of them are based on 3D median or average filters, which supports are along motion trajectories. In this approach, it is difficult to design the proper structure of the 3D filter by analytic methods. The artificial neural networks can be useful tool for creating the structures of the filters. In this paper the novel neural networks approach to motion compensated temporal and spatio-temporal filtering is proposed. The multilayer perceptrons and functional-link nets are used for the 3D filtering. The spatio-temporal patterns are creating from real motion video images. the neural networks learn these patterns. The practical examples of the filtering are shown and compared with traditional motion-compensated filters.

This paper presents a framework for integrating multiple sensory data, sparse range data and dense depth maps from shape from shading in order to improve the 3D reconstruction of visible surfaces of 3D objects. The integration process is based on propagating the error difference between the two data sets by fitting a surface to that difference and using it to correct the visible surface obtained from shape from shading. A feedforward neural network is used to fit a surface to the sparse data. We also study the use of the extended Kalman filter for supervised learning and compare it with the backpropagation algorithm. A performance analysis is done to obtain the best neural network architecture and learning algorithm. It is found that the integration of sparse depth measurements has greatly enhanced the 3D visible surface obtained from shape from shading in terms of metric measurements.

Whenever the input training class patterns applied to a one- layered, hard-limited perceptron (OHP) satisfy a certain positive-linear-independence (PLI) condition, the learning of these patterns by the neural network can be done non- iteratively in a few algebraic steps and the recognition of the untrained test patterns can be very accurate and very robust if a special learning scheme - automatic feature extraction - is adopted in the learning mode. In this paper, we report the theoretical foundation, the simplified design analysis of this novel pattern recognition system, and the experiments we carried out with this novel system. The experimental result shows that the learning of four digitized training patterns is close to real-time, and the recognition of the untrained patterns is above 90 percent correct. The ultra-fast learning speed here is due to the non-iterative nature of the novel learning scheme we used in OHP. The high robustness in recognition here is due to the automatic feature extraction scheme we use in the learning mode.

In this paper ,a new method for edge detection in multispectral imags is presented. It is based on the use of the Self-Organizing Map (SOM), Peano scan and a conventional edge detector. The method presented in this paper order the vectors of the original image in such a way that vectors that are near each other according to some similarity criterium have scalar ordering values near each other. This is achieved using a 2D self-organizing map and the Peano scan. After ordering, the original vector image reduces to a gray-value image, and a conventional edge detector can be applied. In this paper, the Laplace and the Canny edge detectors are used. It is shown, that using the proposed method sit is possible to find the same relevant edges that R-ordering based methods find. Furthermore, it is also possible to find edges in images which consist of metameric colors, i.e. images in which every pixel vector maps into the same location in RGB space. This is not possible using conventional edge detectors which use an RGB image as input. Finally, the new method is tested with a real-world airplane image, giving results comparable with R-ordering based methods.

Curvilinear component analysis (CCA) is performed by an original self-organized neural network, which provides a convenient approach for dimension reduction and data exploration. It consists in a non-linear, preserving distances projection of a set of quantizing vectors describing the input space. The CCA technique is applied to the analysis of CGR fractal images of DNA sequences from different species. The CGR method produces images where pixels represent frequency of small sequences of bases revealing nested patterns in DNA sequences.

Content-based Image Retrieval is an area of growing interest. Various approaches exist which use color, texture, and shape for retrieving 'similar' images from a database. However, what do we mean by 'similar'. Traditionally, similarity is interpreted as distance in feature space. But this does not necessarily match the human users' expectations. We report on two human studies, which asked volunteers to select which imags they considered to be 'most like' each image from the Brodatz dataset. Although the images from the Brodatz set have the advantage of being an agreed standard in texture analysis, Brodatz certainly did not select his images with this in mind. The results from this study provide a justification for selecting a subset of the Brodatz data set for use in evaluating texture-based retrieval techniques. Images which humans have difficulty in agreeing which other images are 'most like' are also poor choices for comparison. Our result indicate which images are most likely to be classified as 'similar' by individual humans and that can also serve to evaluate computer-based retrieval techniques.

The majority of scientific papers focusing on wood classification for pencil manufacturing take into account defects and visual appearance. Traditional methodologies are base don texture analysis by co-occurrence matrix, by image modeling, or by tonal measures over the plate surface. In this work, we propose to classify plates of wood without biological defects like insect holes, nodes, and cracks, by analyzing their texture. By this methodology we divide the plate image in several rectangular windows or local areas and reduce the number of gray levels. From each local area, we compute the histogram of difference sand extract texture features, given them as input to a Local Neuro-Fuzzy Network. Those features are from the histogram of differences instead of the image pixels due to their better performance and illumination independence. Among several features like media, contrast, second moment, entropy, and IDN, the last three ones have showed better results for network training. Each LNN output is taken as input to a Partial Neuro-Fuzzy Network (PNFN) classifying a pencil region on the plate. At last, the outputs from the PNFN are taken as input to a Global Fuzzy Logic doing the plate classification. Each pencil classification within the plate is done taking into account each quality index.

In this paper, we propose a method of Shape from Texture. There are some major approaches to estimate 3D shape. Our method uses the peak frequency for feature of texture, as it is known to be used in human perception. Our proposed system is composed of 1D and 2D system. The 1D system estimates a local peak frequency is composed of two steps. First is the feature extraction step. For extracting the feature of image, we use 16 Gabor filters with successive Gauss filters as post smoothing filter. Second is the estimation of a local peak frequency step with neural network. A local peak frequency is estimated from the neural network. We use a three layer network whose parameters are determined by Back Propagation network training. By using neural network, the performance of 16 Gabor filters is demonstrated as efficient as that with more filters. We use this algorithm for separate orientation channel. 2D system inhibits estimated local peak frequency of 4 orientations in 1D system. And we estimate 3D shape from the ratio of local peak frequencies in 2D. This system is not effective for the estimation near object edge. Then, we use Edge information for improving the method.

We are now developing a brain computer with algorithm acquisition function, where a two-level structure is introduced to connect pattern with (meta-)symbol, because we know how to realize algorithm acquisition on symbols. At Level 1 we use a conventional learning method on neural networks, but, at Level 2, we develop a new learning algorithm AST, where an automation-like algorithm with a neural network learning is introduced. This is powerful enough to realize an automatic algorithm acquisition. We will state a two-level structure and the AST learning algorithm. We focus on real-time image understanding which is a realization of human brain with eyes. We will summarize the features of our developing artificial brain system as follows: 1) System for meta-symbol as well as pattern, 2) Architecture artificial memory model to satisfy the features of 1)-3), We introduce a two-level architecture, where the meta-symbol is introduced at Level 2 while the pattern is used for Level 1 as usual.

Detection of contours in biomedical imags is quite often an a priori step to quantification. Using computer facilities, it is now straightforward for a medical expert to draw boundaries around regions of interest. However, accuracy of drawing is an issue, which is rarely addressed although it may be a crucial point when for example one looks for local evolution of boundaries on a series of images. The aim of our study is to correlate the local accuracy of experts' outlines with local features of the underlying image to allow meaningful comparisons of boundaries. Local variability of experts' outlines has been characterized by deriving a set of distances between outlines repeatedly drawn on the same image. Local features of underlying images were extracted from 64 by 64 pixel windows. We have used a two-stage neural network approach in order to deal with complexity of data within windows and to correlate their features with local variability of outlines. Our method has been applied to the quantification of the progression of the Cytomegalovirus infection as observed from a series of retinal angiograms in patients with AIDS. Reconstruction of new windows from the set of primitives obtained from the GHA network shows that the method preserves desired features. Accuracy of the border of infection is properly predicted and allows to generate confidence envelope around every hand-outlined.

Pulse Coupled Neural Networks have been extended and modified to suit image segmentation applications. Previous research demonstrated the ability of a PCNN to ignore noisy variations in intensity and small spatial discontinuities in images that prove beneficial to image segmentation and image smoothing. This paper describes four research and development projects that relate to PCNN segmentation - three different signal processing applications and a CMOS integrated circuit implementation. The software for the diagnosis of Pulmonary Embolism from VQ lung scans uses PCNN in single burst mode for segmenting perfusion and ventilation images. The second project is attempting to detect ischemia by comparing 3D SPECT images of the heart obtained during stress and rest conditions, respectively. The third application is a space science project which deals with the study of global aurora images obtained from UV Imager. The paper also describes the hardware implementation of PCNN algorithm as an electro-optical chip.

The project consists in extracting biological objects from the background of an image in order to determine their three dimensions, namely their thickness. The small size of the photographed objects induces the formation of light interferences. The observed interference colors are related to the properties of the thin objects. Segmentation techniques used for this application are divided into three major types: edge extraction, region growing and splitting, clustering. Generally, edge segmentation works on each separated RGB channel but it leads to a data fusion problem Region growing and splitting methods commonly deal with features extraction. Color is a possible feature. The color image segmentation can be either monodimensional or multidimensional, using classification methods. For the monodimensional segmentation, the gray level is used alone. For the multidimensional case, one can take into account the vectorial character of colors, using color clustering. In this general context the aim of the project is to evaluate how a specific color space can improve the segmentation. Standard color segmentation algorithms are used: (1) C- means; (2) Back-propagation neural network; (3) Learning Vector Quantization. The results are compared with gray level algorithms such as the Otsu thresholding and ISODATA. Applied to each color channel. They show first that there is not only one good color representation space, and secondly, that data clusters are relatively close to each other, which explains why segmentation is so difficult in this class of pictures.

In the present paper, a new synthesis approach is developed for associate memories based on a modified relaxation algorithm. The design problem, of feedback neural networks for associative memories is formulated as a set of linear inequalities such that the use of pseudo relaxation method is evident. The pseudo relaxation training in the synthesis algorithms is guaranteed to converge for the design of neural networks without any constraints on the connection matrix. To demonstrate the applicability of the present result and to compare the present synthesis approach with existing design methods, a pattern recognition example is considered.