Algorithms of filtering, edge detection, and extraction of details and their implementation using cellular neural networks (CNN) are developed in this paper. The theory of CNN based on universal binary neurons (UBN) is also developed. A new learning algorithm for this type of neurons is carried out. Implementation of low-pass filtering algorithms using CNN is considered. Separate processing of the binary planes of gray-scale images is proposed. Algorithms of edge detection and impulsive noise filtering based on this approach and their implementation using CNN-UBN are presented. Algorithms of frequency correction reduced to filtering in the spatial domain are considered. These algorithms make it possible to extract details of given sizes. Implementation of such algorithms using CNN is presented. Finally, a general strategy of gray-scale image processing using CNN is considered.
The problem of histogram thresholding is tackled using a modular expert network. The modular expert network is a network of expert modules modulated by a gating network. The expert modules incorporate individual experts' opinions on the thresholding problem. The difficult task of integration of conflicting experts' opinions is achieved through a training of the gating network using backpropagation. The resulting network achieves accurate modeling of the solution mapping through the efficient combination of existing experts. Experimental results show the superior performance of the
modular network over classical algorithms. In particular, a near-optimal solution was shown to be achievable using a small training set. Application to a real-world biomedical cell segmentation problem is also given.
This paper describes a novel neural network (NN) based system for detecting rigid objects from their (2-D) gray-level image images. In this approach a labeled graph is employed to construct a template model for an object from the image region where the object is located. A novel network of NN (NoNN) is proposed to learn the examples of object model graph templates. The NoNN is composed of a set of subnetworks that are not connected to one another. The selected network topology improves the generalization of the classifier in terms of its Vapnik-Chervonenkis dimension (VCdim). Each subnetwork is a network of multilayer perceptron neural network
classifiers operating in parallel with the rest of the system. Each subnetwork is assigned to learn the label of one vertex of a graph. The detection scheme combines the decisions of the subnetworks to classify an image graph extracted from an input image block. This visual computational model is potentially useful for partial matching where the object is occluded. Performance of the system is tested in modeling and detection of human eye regions in face images with some degree of variation in the direction of pose.
Character recognition systems can contribute tremendously to the advancement of the automation process and can improve the interaction between person and machine in many applications, including office automation, check verification, and a large variety of banking, business, and data entry applications. Our main theme is the automatic recognition of hand-printed Latin characters using artificial neural networks in combination with conventional techniques. This approach has a number of advantages: it combines rule-based (structural) and classification tests; it is more efficient for large and complex sets; and feature extraction is inexpensive and execution time is independent of handwriting style and size. The technique can be divided into three major steps. The first step is preprocessing in which the original image is transformed into a binary image utilizing a 300 dpi scanner and then thinned using a parallel thinning algorithm. Second, the image skeleton is traced from left to right to build a binary tree. Some primitives, such as straight lines, curves, and loops are extracted from the binary tree. Finally, a three layer artificial neural network is used for character classification. The system was tested on a sample of handwritten characters from several individuals whose writing ranged from acceptable to poor in quality and the correct recognition rate obtained was 91%.
We present some preliminary study results of an automated fingerprint pattern classification approach based on a novel neural network structure called the fuzzy cerebellar model arithmetic computer (CMAC) neural network. The fingerprint images are first preprocessed to generate ridge flow, then the Karhunen-Loever (K-L) transform is used to extract the features from the ridge-flow images. The feature vector is then sent to a fuzzy CMAC neural network for classification. Excellent results were obtained through our preliminary experiments on the "two classes" problem.
This paper introduces a neuro-fuzzy system for the estimation of the crowding level in a scene. Monitoring the number of people present in a given indoor environment is a requirement in a variety of surveillance applications. In the present work, crowding has to be estimated from the image processing of visual scenes collected via a TV camera. A suitable preprocessing of the images, along with an ad hoc feature extraction process, is discussed. Estimation of the crowding level in the feature space is described in terms of a fuzzy decision rule, which relies on the membership of input patterns to a set of partially overlapping crowding classes, comprehensive of doubt classifications and outliers. A society of neural networks, either multilayer perceptrons or hyper radial basis functions, is trained to model individual class-membership functions. Integration of the neural nets within the fuzzy decision rule results in an overall neuro-fuzzy classifier. Important topics concerning the generalization ability, the robustness, the adaptivity and the performance evaluation of the system are explored. Experiments with
real-world data were accomplished, comparing the present approach with statistical pattern recognition techniques, namely linear discriminant analysis and nearest neighbor. Experimental results validate the neuro-fuzzy approach to a large extent. The system is currently working successfully as a part of a monitoring system in the Dinegro underground station in Genoa, Italy.
A new method of texture classification comprising two processing stages, namely a low-level evolutionary feature extraction based on Gabor wavelets and a high-level neural network based pattern recognition, is proposed. The design of these stages is motivated by the processes involved in the human visual system: low-level receptors responsible for early vision processing and the high-level cognition. Gabor wavelets are used as extractors of "lowlevel" features that feed the feature-adaptive adaptive resonance theory (ART) neural network acting as a high-level "cognitive system." The novelty of the model developed in this paper lies in the use of a self-organizing input layer to the fuzzy ART. Evaluation of the model is performed by using natural textures, and results obtained show that the developed model is capable of performing the texture recognition task effectively. Applications of the developed model include the study of artificial vision systems motivated by the human visual system model.
A spectral classification comparison was performed using four different classifiers, the parametric maximum likelihood classifier and three nonparametric classifiers: neural networks, fuzzy rules, and fuzzy neural networks. The input image data is a System Pour l'Observation de la Terre (SPOT) satellite image of Otago Harbour near Dunedin, New Zealand. The SPOT image data contains three spectral bands in the green, red, and visible infrared portions of the electromagnetic spectrum. The specific area contains intertidal vegetation species above and below the waterline. Of specific interest is eelgrass (Zostera novazelandica), which is a biotic indicator of environmental health. The mixed covertypes observed in an in situ field survey are difficult to classify because of subjectivity and water's preferential absorption of the visible infrared spectrum. In this analysis, each of the classifiers were applied to the data in two different testing procedures. In the first test procedure, the reference data was divided into training and test by area. Although this is an efficient data handling technique, the classifier is not presented with all of the subtle microclimate variations. In the second test procedure, the same reference areas were amalgamated and randomly sorted into training and test data. The amalgamation and sorting were performed external to the analysis software. For the first testing procedure, the highest testing accuracy was obtained through the use of fuzzy inferences at 89%. In the second testing procedure, the maximum likelihood classifier and the fuzzy neural networks provided the best results. Although the testing accuracy for the maximum likelihood classifier and the fuzzy neural networks were similar, the latter algorithm has additional features, such as rules extraction, explanation, and fine tuning of individual classes.
A modified error diffusion algorithm that results in a homogenous pulse distribution in the highlight and shadow areas of the binarized images is described. The distinction to standard error diffusion algorithms is a dynamic threshold imprint function that depends on the local input values and binary output pixel. The imprint function is generated by diffusing a one-dimensional function in the orthogonal direction, thereby allowing a fast implementation of a two-dimensional threshold imprint.
The error diffusion halftoning method preserves details well, but produces some unwanted regular texture patterns. The purpose of this paper is to introduce certain nonlinear operators with small kernels for the error diffusion to reduce the regular patterns. The goal is to suppress pattern artifacts while maintaining a small neighborhood. The method employed uses nonlinear diffusion operators, which possess a relatively complex distribution mechanism,
thereby suppressing noticeable patterns. Two nonlinear filter classes are considered: polynomial and median type filters. We found that reduction of regular patterns without producing excessively grainy images is obtained using a combination of linear and median error feedback operators.
One important task in the field of digital video signal processing is the conversion of one standard into another with different field and scan rates. Therefore a new vector-based nonlinear upconversion algorithm has been developed that applies nonlinear center weighted median filters (CWM). Assuming a two channel model of the human visual system with different spatio-temporal characteristics, there are contrary demands for the CWM filters. One can meet these demands by a vertical band separation and an application of so-called temporally and spatially dominated CWMs. By this means, interpolation errors of the separated channels can be compensated by an adequate splitting of the spectrum. Therefore a very robust vector error tolerant upconversion method can be achieved, which significantly improves the interpolation quality. By an appropriate choice of the CWM filter root structures, main picture elements are interpolated correctly even if faulty vector fields occur. To demonstrate the correctness of the deduced interpolation scheme, picture content is classified. These classes are distinguished by correct or incorrect vector assignment and correlated or noncorrelated picture content. The mode of operation of the new algorithm is portrayed for each class. Whereas the mode of operation for correlated picture content can be shown by object models, this is shown for noncorrelated picture content by the probability distribution function of the applied CWM filters. The new algorithm has been verified by objective evaluation methods [peak signal to noise ratio (PSNR), and subjective mean square error (SMSE) measurements] and by a comprehensive subjective test series.
A simple dynamic model of a neural network is presented. Using the dynamic model of a neural network, we improve the performance of a three-layer multilayer perceptron (MLP). The dynamic model of a MLP is used to make fundamental changes in the network optimization strategy. These changes are: Neuron activation functions are used, which reduce the probability of singular Jacobians; Successive regularization is used to constrain the volume of the weight space being minimized; Boltzmann pruning is used to constrain the dimension of the weight space; and prior class probabilities are used to normalize all error calculations, so that statistically significant samples of rare but important classes can be included without distortion of the error surface. All four of these changes are made in the inner loop of a conjugate gradient optimization iteration and are intended to simplify the training dynamics ofthe optimization. On handprinted digits and fingerprint classification problems, these modifications improve error-reject performance by factors between 2 and 4 and reduce network size by 40 to 60%.