An optical modeless Sign Language Recognition (SLR) system is presented. The system uses the HAusdorf-Voronoi NETwork (HAVNET), an artificial neural network designed for 2D binary pattern recognition. It uses adaptation of the Hausdorff distance to determine the similarity between an input pattern and a learned representation. A detailed review of the architecture, the learning equations, and the recognition equations for the HAVNET network are presented. Competitive learning has been implemented in training the network using a nearest-neighbor technique. The SLR system is applied to the optical recognition of 24 static symbols from the American Sign Language convention. The SLR system represents the target images in a 80 X 80 pixel format. The implemented HAVNET network classifies the inputs into categories representing each of the symbols, using an output layer of 24 nodes. The network is trained with 5 different formats for each symbol and is tested with all 24 symbols in 15 new formats. Results from the SLR system without competitive training show shape identification problems, when distinguishing symbols with similar shapes. Implementation of competitive learning in the HAVNET neural network improved recognition accuracy on this task to 89%. The hand gestures are identified through a window search algorithm. Feature recognition is obtained from edge enhancement by applying a Laplacian filter and thresholding, which provides robustness to pose, color and background variations.

This work is in the field of automated document processing. This work addresses the problem of representation and recognition of Urdu characters based on concepts from Fuzzy logic. In particular, we show that Fuzzy logic is used here to make a classification of 36 Urdu characters into seven sub-classes namely sub-classes characterized by seven proposed and defined fuzzy features specifically related to Urdu characters. We show that here Fuzzy logic provides a remarkably simple way to draw definite conclusions from vague, ambiguous, noisy or imprecise information. In particular, we illustrate the concept of `interest regions' and describe a framing method that provides a way to make the proposed technique for Urdu characters recognition robust and invariant to scaling and translation. We also show that a given character rotation is dealt with by using the Hotelling transform. This transform is based upon the eigenvalue decomposition of the covariance matrix of an image, providing a method of determining the orientation of the major axis of an object within an image. Finally experimental results are presented to show the power and robustness of the proposed Fuzzy logic based technique for Urdu character recognition, its fault tolerance, and high recognition accuracy.

In this paper, we present a new system to segment and label document images into text, halftone images, and background using feature extraction and unsupervised clustering. Each pixel is assigned a feature pattern consisting of a scaled family of differential geometrical invariant features and texture features extracted from the cooccurence matrix. The invariant feature pattern is then assigned to a specific region using a two-stage neural network system. The first stage is a self-organizing principal components analysis (SOPCA) network that is used to project the feature vector onto its leading principal axes found by using principal components analysis. Using the SOPCA algorithm, we can train the SOPCA network to project our feature vector orthogonally onto the subspace spanned by the eigenvectors belonging to the largest eigenvalues. By doing that we ensure that the vector is represented by a reduced number of effective features. The next step is to cluster the output of the SOPCA network into different regions. This is accomplished using a self-organizing feature-map (SOFM) network. In this paper, we demonstrate the power of the SOPCA-SOFM approach to segment document images into text, halftone, and background.

This paper presents a new lossless image compression technique called modular differential pulse code modulation. The proposed technique consists of a VQ classifier and several neural network class predictors. The classifier uses the four previously encoded pixels to identify the class of the current pixel (the pixel to be predicted). The current pixel is then predicted by the corresponding class predictor. Experimental results demonstrate that the proposed technique reduces the bit rate by as much as 10 percent when compared to the lossless JPEG.

This paper deals with application of fuzzy and neural techniques to the reversible intraframe compression of grayscale images. With reference to a spatial DPCM scheme, prediction may be accomplished in a space varying fashion following two main strategies: adaptive, i.e., with predictors recalculated at each pixel position, and classified, in which image blocks, or pixels are preliminarily labeled into a number of statistical classes, for which minimum MSE predictors are calculated. Here, a trade off between the above two strategies is proposed, which relies on a space-varying linear-regression prediction obtained through fuzzy techniques, and is followed by context based statistical modeling of predictive errors, to enhance entropy coding. A thorough comparison with the most advanced methods in the literature, as well as an investigation of performance trends to work parameters, highlight the advantages of the fuzzy approach.

The paper presents an application of data mining methods for face recognition. We have proposed the rough sets methods to face recognition and selection of facial features based on the minimum concept description paradigm. The features from the face images have been extracted based on the singular value decomposition, followed by the principal component analysis, and rough sets processing. The recognition of facial images, for reduced feature sets, has been carried on using rough sets expert system and error backpropagation neural network.

Prediction of landcover type from airborne/spaceborne sensors is an important classification problem in remote sensing. Due to advances in sensor technology, it is now possible to acquire hyperspectral data simultaneously in more than 100 bands, each of which measures the integrated response of a target over a narrow window of the electromagnetic spectrum. The bands are ordered by their wavelengths and spectrally adjacent bands are generally statistically correlated. Using such high dimensional data for classification of landcover potentially provides greatly improved results. However, it is necessary to select bands that provide the best possible discrimination among the classes of interest. In this paper, we propose an efficient top-down multiresolution class-dependent feature extraction algorithm for hyperspectral data to be used with a pairwise classification scheme. First, the C class problem is divided into (^C₂) two class problems. Features for each pair of classes are extracted independently. The algorithm decomposes the bands recursively into groups of adjacent bands (subspaces) in a top-down fashion. The features extracted are specific to the pair of classes that are being distinguished and exploit the ordering information in the hyperspectral data. Experiments on a 183 band AVIRIS data set for a 12 class problem show significant improvements in both classification accuracies and the number of features required for all 66 pairs of classes.

Due to the proliferation of sensor-platform combinations that are capable of wide area searches, automated target detection has become increasingly important. Most learning- algorithm-based target detectors perform dimensionality reduction before actual training, because the high dimensionality of imagery requires enormous training sets to achieve satisfactory performance. One potential problem with this approach is that most dimensionality reduction techniques, such as principal component analysis, seek to maximize the representation of data variation into each component in turn, without considering interclass discriminability. We present a neural-network-based transformation that provides dimensionality reduction and a high degree of discriminability. Our approach achieves simultaneous data compression and target discriminability by adjusting the pretrained base components to maximize separability between classes. This will allow classifiers to operate at a higher level of efficiency and generalization capability on the low-dimensionality data that contain highly discriminative information as well.

In this paper, we propose an evolutionary neural network for blind source separation (BSS). The BSS is the problem to obtain the independent components of original source signals from mixed signals. The original sources that are mutually independent and are mixed linearly by an unknown matrix are retrieved by a separating procedure based on Independent Component Analysis (ICA). The goal of ICA is to find a separating matrix so that the separated signals are as independent as possible. In neural realizations, separating matrix is represented as connection weights of networks and usually updated by learning formulae. The effectiveness of the algorithms, however, is affected by the neuron activation functions that depend on the probability distribution of the signals. In our method, the network is evolved by Genetic Algorithm (GA) that does not need activation functions and works on evolutionary mechanism. The kurtosis that is a simple and original criterion for independence is used in the fitness function of GA. After learning, the network can be used to separate other mixed signals of the same mixing procedure. The applicability of the proposed method for blind source separation is demonstrated by the simulation results.

Classification of single particle projection images of heterogeneous sets before 2D and 3D analysis is still a major problem in electron microscopy. Images obtained by the microscope not only present a very low signal/noise ratio but also a wide range of variability due to the non homogeneous background on which particles lay and tilting differences among other factors. Blind classification procedures are therefore bound to fail or in any case can be hardly reliable, thus making necessary the use of dimensionality reduction tools in order to ease the task of classification and to introduce some kind of control over the process. The purpose of this work is the evaluation of both linear and nonlinear unsupervised feature extraction techniques together with several pattern recognition and automatic classification tools, some of which have not yet been applied and tested in this context. Mapping and classification procedures include statistical and neural network tools.

This paper presents a novel technique for texture modeling and synthesis using the random neural network (RNN). This technique is based on learning the weights of a recurrent network directly from the texture image. The same trained recurrent network is then used to generate a synthetic texture that imitates the original one. The proposed texture learning technique is very efficient and its computation time is much smaller than that of approaches using Markov Random Fields. Texture generation is also very fast. We have tested our method with different synthetic and natural textures. The experimental results show that the RNN can efficiently model a large category of homogeneous microtextures. Statistical features extracted from the co- occurrence matrix of the original and the RNN based texture are used to evaluate the quality of fit of the RNN based approach.

The problem of color image enhancement and the specific case of color demosaicing which involves reconstruction of color images from sampled images, is an under-constrained problem. Using single-channel restoration techniques on each color- channel separately results in poorly reconstructed images. It has been shown that better results can be obtained by considering the cross-channel correlation. In this paper, a novel approach to demosaicing is presented, using learning schemes based on Artificial Neural Networks. Thus the reconstruction parameters are determined specifically for predefined classes of images. This approach improves results for images of the learned class, since the variability of inputs is constrained (within the image class) and the parameters are robust due to the learning process. Three reconstruction methods are presented in this work. Additionally, a selection method is introduced, which combines several reconstruction methods and applies the best method for each input.

The paper describes a neural-based method for matching spatially distorted image sets. The matching of partially overlapping images is important in many applications-- integrating information from images formed from different spectral ranges, detecting changes in a scene and identifying objects of differing orientations and sizes. Our approach consists of extracting contour features from both images, describing the contour curves as sets of line segments, comparing these sets, determining the corresponding curves and their common reference points, calculating the image-to-image co-ordinate transformation parameters on the basis of the most successful variant of the derived curve relationships. The main steps are performed by custom neural networks. The algorithms describe in this paper have been successfully tested on a large set of images of the same terrain taken in different spectral ranges, at different seasons and rotated by various angles. In general, this experimental verification indicates that the proposed method for image fusion allows the robust detection of similar objects in noisy, distorted scenes where traditional approaches often fail.

Hubble's pioneering discovery of the distance to the Andromeda galaxy opened the frontiers of extragalactic research. According to estimates computed from the Hubble Deep Field, astronomers predict that the universe may potentially contain over 50 billion galaxies. Recognition/Classification of galaxies is an important issue in the large-scale study of the Universe and is not a simple task. Several techniques have been reported for the classification of galaxies. Artificial neural networks are being successfully used for classification in various applications. Recently, the Pulse-Coupled Neural Network (PCNN) has been shown to be useful for image pre-processing. When a digital image is applied to the input of a PCNN, the network groups image pixels based on spatial proximity and brightness similarity. A time signal can be obtained by computing the number of `on pixels' for each iteration; the time signal being an encoded 1D signature of a 2D image. There is a one-to-one correspondence between images and their time signatures. In the current study, we exploit this property to obtain time signatures of optical galaxies. Our results on spirals, ellipticals and irregulars are very promising and the work is being extended towards the development of an efficient and robust computer automated classifier.

Individual region-of-interest atlas extraction consists of two main parts: T₁-weighted MRI grayscale images are classified into brain tissues types (gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), scalp/bone (SB), background (BG)), followed by class image analysis to define automatically meaningful ROIs (e.g., cerebellum, cerebral lobes, etc.). The purpose of this algorithm is the automatic detection of training points for neural network-based classification of brain tissue types. One transaxial slice of the patient data set is analyzed. Background separation is done by simple region growing. A random generator extracts spatially uniformly distributed training points of class BG from that region. For WM training point extraction (TPE), the homogeneity operator is the most important. The most homogeneous voxels define the region for WM TPE. They are extracted by analyzing the cumulative histogram of the homogeneity operator response. Assuming a Gaussian gray value distribution in WM, a random number is used as a probabilistic threshold for TPE. Similarly, non-white matter and non-background regions are analyzed for GM and CSF training points. For SB TPE, the distance from the BG region is an additional feature. Simulated and real 3D MRI images are analyzed and error rates for TPE and classification calculated.

In this paper, we do some pre-processing on the input data to remove some noise before putting them into the network and some post-processing before outputting the results. Different neural networks such as back-propagation, radias basis network with different architecture are tested. We choose the one with the best performance among them. From the experiments we can see that the results of the neural network are similar to those given by the experienced doctors and better than those of previous research, indicating that this approach is very practical and beneficial to doctors comparing with some other methods currently existing.

A shape-based classification method is developed based upon the Generalized Fourier Representation (GFR). GFR can be regarded as an extension of traditional polar Fourier descriptors, suitable for description of closed objects, both convex and concave, with or without holes. Explicit relations of GFR coefficients to regular moments, moment invariants and affine moment invariants are given in the paper. The dual linear relation between GFR coefficients and regular moments was used to compare shape features derive from GFR descriptors and Hu's moment invariants. the GFR was then applied to a clinical problem within oral medicine and used to represent the contours of the lesions in the oral cavity. The lesions studied were leukoplakia and different forms of lichenoid reactions. Shape features were extracted from GFR coefficients in order to classify potentially cancerous oral lesions. Alternative classifiers were investigated based on a multilayer perceptron with different architectures and extensions. The overall classification accuracy for recognition of potentially cancerous oral lesions when using neural network classifier was 85%, while the classification between leukoplakia and reticular lichenoid reactions gave 96% (5-fold cross-validated) recognition rate.

A Tongue Imaging and Analysis System is being developed to acquire digital color tongue images, and to automatically classify and quantify the tongue characteristics for traditional Chinese medical examinations. An important processing step is to segment the tongue pixels into two categories, the tongue body (no coating) and the coating. In this paper, we present a two-stage clustering algorithm that combines Fuzzy Kohonen Clustering Networks and Genetic Algorithm for the segmentation, of which the major concern is to increase the interclass distance and at the same time decrease the intraclass distance. Experimental results confirm the effectiveness of this algorithm.

A novel scheme using spherical manifolds is proposed for the simultaneous classification and pose estimation of 3D objects from 2D images. The spherical manifold imposes a local topological constraint on samples that are close to each other, while maintaining a global structure. Each node on the spherical manifold also corresponds nicely to a pose on a viewing sphere with 2 degrees of freedom. The proposed system is applied to aircraft classification and pose estimation.

The detection of image segmented objects in video sequences is constrained by the a priori information available with a classifier. An object recognizer labels image regions based on texture and shape information about objects for which historical data is available. The introduction of a new object would culminate in its misclassification as the closest possible object known to the recognizer. Neural networks can be used to develop a strategy to automatically recognize new objects in image scenes that can be separated from other data for manual labeling. In this paper, one such strategy is presented for natural scene analysis of FLIR images. Appropriate threshold tests for classification are developed for separating known from unknown information. The results show that very high success rates can be obtained using neural networks for the labeling of new objects in scene analysis.

There is a considerably history of using Eddy Current measurements for inspecting metal surfaces. The interpretation of the results, however, has been somewhat subjective. The sensitivity of the Eddy Current measurement to changes in material thickness is easily demonstrated. However, the reconstruction of thickness from the measurements (with any degree of confidence) has remained elusive. Part of the issue is the additional sensitivity of Eddy Current measurements to Lift distance and local geometry. This project is a culmination of an feasibility study to see if neural networks could provide estimates of material thicknesses for thin materials (< 0.15 inches or 4 mm) for metals with low conductivity (approximately 0.7 x mega/Ohm cm). The study a neural network model, whose accuracy varies depending upon the conditions. It turns out that the conditions required for good accuracy are usually not the conditions one wants to take measurements. The model not only suggests why this problem has been unwieldy, but also suggests that most of the difficulties could be alleviated by using an independent source to measure the Lift distance.

In this paper, a new online inspection method for part configuration and surface is presented by combining computer vision and neural networks. Different from conventional contact measurement, it is non-contact measurement method, and it can operate on-line. In this method, the 3D configuration and surface of part are reconstructed from stereo image pair taken by computer vision system. The architecture for parallel implementation of part measurement system is developed using neural networks. Several relevant approaches including system calibration, stereo matching, and 3D reconstruction are constructed using neural networks. Instead of conventional system calibration method that needs complicated iteration calculation process, the new system calibration approach is presented using BP neural network.

This paper describes a full-custom mixed-signal chip which embeds distributed optical signal acquisition, digitally- programmable analog parallel processing, and distributed image memory--cache--on a common silicon substrate. The paper briefly describes the chip architecture and focus mostly on presenting experimental evidence of the chip functionality. Multiscale low-pass and high-pass filtering of gray-scale images, analog edges extraction, image segmentation, thresholded gradient detection, mathematical morphology operations, shortest path detection in a labyrinth, skeletonizing, image reconstruction, several non- linear type image processing tasks like absolute value calculation of gray-scale gradient detection and real-time motion detection in QCIF video sequences are some of the very interesting applications that have been demonstrated as available when using the prototype.

In this paper, we realize the classification of the gray cast iron according to the graphite morphology in it by Artificial Neural Network. It's a part of a big metallurgic analytical software system, and also takes on some significance in the automatic production in iron and steel industry. Our work is described as 2 steps here: The first one is texture feature extracting and the second one, classification. The images we worked on come from metallographic electron microscope, and in needs, we do some pretreatment on it. The textural features extracted mainly based on fractal parameter, roughness parameter and regression, and some comparison is also made between these textural modes. The classification is performed through artificial neural network--multilayer back-propagation neural network, which is based on a kind of feed-forward artificial neural network. It learns samples and trains itself by BP algorithm--error back propagation algorithm. To reduce the computational quantity, we obtain the number of hidden nodes directly by the numbers of input nodes and output nodes. Result shows available.