In this paper, an adaptive neural network vector predictor is designed in order to improve the performance of the predictive component of the predictive vector quantizer (PVQ). The proposed vector predictor consists of a set of dedicated predictors (experts) where each predictor is optimized for a particular class of input vectors. In our simulations, we used five multi-layer perceptrons (MLP) to design our expert predictors. Each MLP predictor is separately trained by using a set of training vectors that belong to a particular class. The class identity of each training vector is determined by its directional variances. In our current implementation, one predictor is optimized for stationary blocks and four other predictors are designed for horizontal, vertical, 45 degree and 135 degree diagonally oriented edge blocks. The back-propagation algorithm is used for training each network. The directional variances of the neighboring blocks are used to select the appropriate expert predictor for the current input block. Therefore, no overhead information is transmitted in order to inform the receiver about the predictor selection. Our simulation shows that the proposed scheme gives an improvement of more than 1 dB over the predictor consisting of a single MLP predictor. The perceptual quality of the predicted images are also significantly improved.

A finite-state auto-associative MLP neural network is studied in the context of dimensionality reduction and smooth signal reconstruction. We describe the structure and the training procedure of the finite-state network. One of the desirable properties of the auto-associative MLP is that the variance of the hidden units' outputs can be arranged in a descending order, so that efficient coding of the hidden layer output can be implemented. We provide experimental results to demonstrate that the finite-state network retains this desirable property of its memory-less counterpart. One of the application areas of the auto-associative MLP is image compression. As with other block based image compression techniques, this method cannot avoid the problem of annoying 'blocking effects' in the reconstructed images. We present simulation results to demonstrate that the finite-state auto-associative MLP can be used to achieve effective image data compression while significantly reducing the blocking effects.

In this paper we propose a new method of image quality assessment for the evaluation of the block distortion through artificial neural networks (ANNs). The approach is new and intends to address the problem of the assessment of the visual quality of compressed images from an original point of view. ANNs in particular are applied in order to detect the presence of blocking errors inside pre-processed pictures. To this purpose, a new local blocking distortion parameter is introduced. Experiments and simulations, even if very preliminary, have confirmed the interest of the proposed approach. A complete formalization of the problem also is presented.

In this paper we present a new scheme for color image compression. The proposed scheme exploits the correlation between the basic color components (red, green, and blue: RGB) by predicting two color components given one color component. Specifically, this scheme employs neural network predictors to predict the red and blue color components using the encoded (reconstructed) green color component. The prediction error is further quantized using vector quantization. The performance of the proposed scheme is evaluated and compared with that of the JPEG.

In this paper, color vector quantization is performed by a competitive learning based clustering algorithm with some modifications that eliminate the false colors that may appear on the resulting image. The preliminary operations that must be applied to the input image pixels before the algorithm can be applied are also stated. Moreover, it is demonstrated that with this scheme, faster convergence and less computations are possible using only a small fraction of all the pixels, but at the same time producing satisfactory results. Finally the results are compared to those of the K-means clustering algorithm.

A practical approach to continuous-tone color image transformation into the image with a finite number of colors is proposed. To achieve this the self organized network was explored. The basic feature of this network provides its self-learning by the competition between several hypotheses (colors) about the analyzed image and the most authentic hypothesis overcomes. Unlike traditional algorithms of color separation which tend to analyze the quantitative contribution of different color components (red, green, blue) to the initial image, the proposed network moreover takes into consideration the qualitative, statistical character of their distribution. To make the operation of self training in learning mode for this network more accurate the output neurons are connected additionally by lateral excitement connections. We proposed the estimate to measure the algorithm efficiency and similarity between input and output images, which gives the possibility to compare our method with other well-known methods. Our method is useful in the fields of precision color image analysis and understanding.

In this note, we present our endeavors to segment same cross-sections of the human brain obtained from the two modalities -- x-ray computed tomography (CT) and magnetic resonance imaging (MRI) -- using the fuzzy c-means technique developed by Bezdek. The two advantages of the technique are that it is unsupervised and is robust to missing and noisy data. Attempts at integrating the images from these two modalities are also mentioned.

Cellular neural networks (CNNs) are currently being used to arrive at solutions to the problems in image processing and pattern recognition. In this paper a technique for image enhancement using a fuzzy CNN is proposed. This technique exploits the massive parallelism of CNNs and mathematical framework of fuzzy logic to cope respectively with the computational complexity and uncertainty in noisy image. The mathematical model of discrete time cellular neural network (DTCNN) is obtained from the circuit equations of a cell. An architecture of fuzzy CNN for image enhancement is proposed. The network extracts the original image from a given noisy image by self organization. The fuzziness of the output image at each iteration is taken as a measure of error, which is in turn used to adapt the input image. An algorithm for adaptation of input image for linear and quadratic indices of fuzziness is derived. The efficacy of the proposed technique is verified through simulation results. Tests are carried out on noisy images obtained by adding zero mean Gaussian noise to synthetic bitonic images. The application of the proposed network for enhancement of noisy images with different noise levels is demonstrated.

This paper describes an application of neural networks in segmenting gray shade images. It describes a method for ranking pixel features relative to their ability to discriminate among different image segment classes. A neural classifier is proposed which operates on pixel feature vectors as inputs to the network, each feature having a variable weight. The weights are iteratively changed to obtain dense and highly separated clusters. The resulting weights are indicative of the usefulness, or rank, of the features.

We are interested in designing a neural network system for automatic chromosome. The goal of this approach is to make the chromosome regions more salient and more interpretable to human skilled technicians than they are in the original imagery. The proposed segmentation model is based upon the biologically derived boundary contour system (BCS) of Grossberg and Mingolla. The practical application of the model to real images raises an important problem. The boundaries generated by BCS have a sizable thickness that is a function of the contrast gradient between two adjacent regions. In order to solve this problem we propose the use of a feedback diffusion. The image resultant of the diffusion is fed back to the simple cell layer. Furthermore, the boundary representation is also fed back to the boundary segmentation stage. In this way, the boundaries are adapted to the variations produced by the feedback diffusion, achieving a gradual boundary thinning. We also propose a modificated diffusive filling-in equation for obtaining better results in homogeneous regions. The behavior of the Grossberg-Todorovic's equation reduces the homogenizing of the regions contained inside the boundaries. In order to solve this problem we introduce a new parameter, rho, called recovery parameter. This parameter regulates the activity variation margin of a node with respect to its initial value. With regard to the improvement in homogenizing, with a value for parameter rho near to zero, the resulting regions present a plain surface, making easy the chromosome bands separation.

Convolution operators act as matched filters for certain types of variations found in images and have been extensively used in the analysis of images. However, filtering through a bank of N filters generates N filtered images, consequently increasing the amount of data considerably. Moreover, not all these filters have the same discriminatory capabilities for the individual images, thus making the task of any classifier difficult. In this paper, we use genetic algorithms to select a subset of relevant filters. Genetic algorithms represent a class of adaptive search techniques where the processes are similar to natural selection of biological evolution. The steady state model (GENITOR) has been used in this paper. The reduction of filters improves the performance of the classifier (which in this paper is the multi-layer perceptron neural network) and furthermore reduces the computational requirement. In this study we use the Laws filters which were proposed for the analysis of texture images. Our aim is to recognize the different textures on the images using the reduced filter set.

A method for representing images in the expanded threshold basis and the space of information vectors is proposed. A method for synthesis of combinational schemes on one neural element and module 2 adders is suggested. The functionals of similarity and difference of p-fragments in expanded threshold basis have been constructed.

We consider the problem of recognition of rigid, manufactured objects, each from a predefined set of possible object classes, from their images. We describe a parametric statistical approach to this problem that is a hybrid between statistical modeling using Bayes decision theory with a generative model of images and a case-based approach. Our method is a variant of the Gibbs sampling method, commonly used to compute posterior probabilities in complex statistical models, but unlike standard Gibbs sampling methods, our method is based directly on analysis of a library of previously analyzed images. We also propose a simple gradient descent method to optimize the parameters of the models to maximize effective object recognition.

In this paper, we report an automatic land cover tracking system which is based on a neural network classifier to extract the land cover from multi-temporal satellite images. The neural network classifier has a three-layer feedforward structure. The input layer has several input units for each of the preprocessed spectral bands of the LANDSAT multispectral scanner, one unit for the digital elevation model, and several units for texture features obtained from a 5 by 5 moving window. The output layer has a neuron for each of the land-cover classes. A pixel is classified with the label of the output layer neuron with the largest activation. The proposed approach provides a quick assessment on the land cover transformation for multitemporal satellite images.

This paper proposes a new hierarchical multilayer perceptron (HMLP) network as an improved classifier for alpha-numeric character recognition from the automobile license plates. HMLP is built up by hierarchically stacking the small MLP subnetworks that work as local classifiers in the feature space. Each subnetwork is trained not to cause overtraining. The generalization and overtraining problem of neural networks was reduced by this method. The classification performance of HMLP was compared with the conventional single stage MLP, and HMLP showed better performance especially for the large sized classification problem.

In this paper, we present a neural network approach for scene analysis: detection of human beings in images. To solve this problem, a precise classification system is required, with adaptation systems based on data processing. These systems must be largely parallel, which is why neural networks have been chosen. The first part of this text is a brief introduction to neural networks and their applications. The second part is a description of the image base composed for experiments and the low-level processing used, then we detail the method used to extract the texture feature of images. The third part describes the Bayesian method and its application to our problem. Part four shows the association of these texture processes with the neural network for identification of human beings. Finally, we conclude with the validity of the method and its future applications.

With feed-forward adaptive network (FFAN) and feed-back associative network (FBAN) respectively imitating the performances of retina and cerebral cortex, an artificial retinal neural network (ARNN) was presented in this paper for fast recognition of visual patterns. In our ARNN model to be implemented with neural network chip MD1200, every synaption of neurons can be arbitrarily given as an integer value from minus 2¹⁵ to 2¹⁵. After these synaptions are trained, the visual pattern not only under geometric transformation but also in the presence of noise can be recognized by the ARNN's system.

A programmable and reconfigurable optical/digital hybrid neural processor with 1024 neurones is described. The relatively large scale and flexibility of the processor benefit by the mutual complement of the optics and an on-line microcomputer. As an example of application, a target classification of 4 kinds of aircraft based on their binary images rotated in a plane by arbitrary angles is performed. The model is a cascaded neural network consisting of three subnets. The principle of the classifier and computer simulation are outlined. In the preliminary experiments, more than 89% of the unlearned images and imperfect ones can be classified correctly.

This paper considers the problem of limited angle tomography in which a complete sinogram is not available. This situation arises in many practical applications where tomographic projection over 180 degrees is either physically unrealizable or infeasible. When a complete sinogram is not available, it is well known that the reconstructed images using common reconstruction algorithms, such as convolution back projection (CBP), will have severe streak artifacts. In this paper, we present a linear artificial neural network to extrapolate the missing part of the sinogram. Once the complete sinogram is obtained via extrapolation, standard reconstruction techniques such as CBP can be used to generate artifact free reconstructions. The parameters of the neural network are designed using the sampling theory of signals with non-compact spectral support, the knowledge that complete sinograms have bowtie-shaped spectral support, and regularization. It is found that once designed, these parameters are data independent, especially for images of similar nature. For sinogram with 2N angular views, each having M raysum per view, if 2L views are available, the computational requirement of the neural network is 4MNL only. Hence, it is much more efficient than other iterative algorithms such as the method of projection onto convex sets, the Papoulis-Gerchberg's algorithm and the Clark-Palmer-Lawrence interpolation method.

A method that makes the Hopfield neural network perform the point pattern relaxation matching process is proposed. An advantage of this is that the relaxation matching process can be performed in real time with the massively parallel capability to process information of the neural network. Experimental results with large simulated images prove the effectiveness and feasibility to perform point relaxation matching by the Hopfield neural network.

In this paper, peak-statistics-based preprocessing methods for detecting point-source target (PST) in IR images are described, and a neural network connection for extracting peak- statistics with recurrent networks is presented. Because of the strong correlation in IR noise, a sample will have less probability in having a larger intensity than its adjacent samples do, in other words, it will have less probability in becoming a peak. The presence of PST interrupts the consistency of correlate between adjacent samples of IR noise, as a result, turning up the difference in peak-statistics. Based on the features of PST and a detailed analysis on the features of IR noise, we adopt a modified difference operation, namely, bidirectional difference (BD), and a peak-statistics-based threshold operation as the preprocessing step. Theory analyses and simulation results have shown that the performance of the proposed methods is better than normal threshold operation.

As we published in the last five years, the supervised learning in a hard-limited perceptron system can be accomplished in a noniterative manner if the input-output mapping to be learned satisfies a certain positive-linear-independency (or PLI) condition. When this condition is satisfied (for most practical pattern recognition applications, this condition should be satisfied,) the connection matrix required to meet this mapping can be obtained noniteratively in one step. Generally, there exist infinitively many solutions for the connection matrix when the PLI condition is satisfied. We can then select an optimum solution such that the recognition of any untrained patterns will become optimally robust in the recognition mode. The learning speed is very fast and close to real-time because the learning process is noniterative and one-step. This paper reports the theoretical analysis and the design of a practical charter recognition system for recognizing hand-written alphabets. The experimental result is recorded in real-time on an unedited video tape for demonstration purposes. It is seen from this real-time movie that the recognition of the untrained hand-written alphabets is invariant to size, location, orientation, and writing sequence, even the training is done with standard size, standard orientation, central location and standard writing sequence.

Traditional forecasting models such as the Box-Jenkins ARIMA model are almost all based on models that assume a linear relationship amongst variables and cannot approximate the non- linear relationship that exists amongst variables in real-world data such as stock-price data. Artificial neural networks, on the other hand, consist of two or more levels of nonlinearity that have been successfully used to approximate the underlying relationships of time series data. Neural networks however, pose a design problem: their optimum topology and training rule parameters including learning rate and momentum, for the problem at hand need to be determined. In this paper, we use genetic algorithms to determine these design parameters. In general genetic algorithms are an optimization method that find solutions to a problem by an evolutionary process based on natural selection. The genetic algorithm searches through the network parameter space and the neural network learning algorithm evaluates the selected parameters. We then use the optimally configured network to predict the stock market price of a blue-chip company on the UK market.