As Internet-enabled computers become ubiquitous in homes, schools, and other publicly accessible locations, there are more people 'surfing the net' who would prefer not to be exposed to offensive material. There is a lot of material freely available on the Internet that we, as a responsible and caring society, would like to keep our children from viewing. Pornographic image content is one category of material over which we would like some control. We have been conducting experiments to determine the effectiveness of using characteristic feature vectors and neural networks to identify semantic image content. This paper will describe our approach to identifying pornographic images using Gabor filters, Principal Component Analysis (PCA), Correllograms, and Neural Networks. In brief, we used a set of 5,000 typical images available from the Internet, 20% of which were judged to be pornographic, to train a neural network. We then apply the trained neural network to feature vectors from images that had not been used in training. We measure our performance as Recall, how many of the verification images labeled pornographic were correctly identified, and Precision, how many images deemed pornographic by the neural network are in fact pornographic. The set of images that were used in the experiment described in this paper for its training and validation sets are freely available on the Internet. Freely available is an important qualifier, since high-resolution, studio-quality pornographic images are often protected by portals that charge members a fee to gain access to their material. Although this is not a hard and fast rule, many of the pornographic images that are available easily and without charge on the Internet are of low image quality. Some of these images are collages or contain textual elements or have had their resolution intentionally lowered to reduce their file size. These are the offensive images that a user, without a credit card, might inadvertently come across on the Internet. Identifying this type of pornographic pictures of low image quality poses particular challenges for any detection software. This paper will address some of the challenges and hurdles we faced in designing and carrying out our experiments. The paper will also discuss the main results of our experiments, as well as some confounds that, at present, limit the effectiveness of our approach to identifying pornographic images, and some directions that may be taken in future research.

Multi-valued neurons (MVN) are the neural processing elements with complex-valued weights and high functionality. It is possible to implement an arbitrary mapping described by partial-defined multiple-valued function on the single MVN. The MVN-based neural networks are applied to temporal classification of images of gene expression patterns, obtained by confocal scanning microscopy. The classification results confirmed the efficiency of this method for image recognition. It was shown that frequency domain of the representation of gene expression images is highly effective for their description.

Many studies for computer-based chromosome analysis using artificial neural network (ANN) have shown that it is possible to classify chromosomes into 24 subgroups. It is important to select optimum features for training the ANN. Our purpose was to select features that had the low classification error and the best ability for human chromosome classification. We applied the medial axis transformation for the medial line, extended the line to the boundary and obtained relative length, relative area and centromeric index. The Giemsa-stained human chromosome has a sequence of banding pattern that is perpendicular to the medial axis of the chromosome. Density profile is a one-dimensional graph of the banding pattern property of the chromosome computed at a sequence of points along the possibly curved chromosome medial axis. Some studied used relative length, centromeric index and 62 density profile as features, but we prepared two data sets as features that one set was relative length, centromeric index and 80 density profile considered No. 1 chromosome's length and the other was relative length, centromeric index, the 80 density profile and relative area and compared classification error of each set. We found that the classification error showed to be decreased by adding relative area to the other features.

This paper presents classification of difference image blocks between the two successive image frames for video data compression. Difference blocks are classified to several activity categories according to the image activity distribution. The classification procedure goes in two steps: activity classification and distribution classification. In the activity classification, each interframe difference image block is classified into active or not-active class according to the amount of motion contained in the block. Distribution classification further classifies active image blocks to four activity categories, vertical, horizontal, diagonal, and uniform activities, based on the activity distribution measured by the edge feature vector in the discrete cosine transform domain. A multiplayer feedforward neural network, trained with a small set of sample classification data, successfully classified difference image blocks according to edge feature distribution. The classification scheme improves the performance of video compression at a cost of small increase in the overhead associated with the quantizer switching.

Most neural networks consisting of discrete (or sign- function) neurons can be studied by discrete mathematics and N-dimension geometry. Particularly, the supervised learning of a feed-forward neural system is crucially related to the geometry of N-dimension convex cones in the N-space. It is shown in this paper that to learn a set of pattern sample vectors forming a convex cone in the N-space, it is only necessary to learn the boundary vectors (or the extreme edges) of this cone, which then makes the learning much more efficient. This paper provides a novel approach to test the convexity of a set of N-vectors (given numerically in an Euclidean N-space) and to find the boundary vectors of this set if it is convex.

Threshold binary networks of the discrete Hopfield-type lead to the efficient retrieval of the regularized least-squares (LS) solution in certain inverse problem formulations. Partitions of these networks are identified based on forms of representation of the data. The objective criterion is optimized using sequential and parallel updates on these partitions. The algorithms consist of minimizing a suboptimal objective criterion in the currently active partition. Once the local minima is attained, an inactive partition is chosen to continue the minimization. This strategy is especially effective when substantial data must be processed by resources which are constrained either in space or available bandwidth.

For a one-layered-feedback neural network, e.g., a Hopfield net, containing discrete sign-function neurons, the nonlinear properties of this network can be studied very efficiently using simple discrete mathematics. This paper summarizes the discrete-formulation of the problem as a matrix difference equation, the simple iterative method of solving this difference equation and the derivation of the major anomalous properties of the system from the solutions. These anomalous properties include, eigen-state storage, associative storage, domain of attraction, content- addressable recall, fault-tolerant recall, capacity of storage, binary oscillating states, limit-cycles in the state space, and noise-sensitive input states. The physical origin and the systematic trend of the derivation of these properties are easily seen in the numerical examples given.

The proposed system for CT image reconstruction is structured with three layers of neurons. In our previous work, we used the resilient backpropagation(Rprop) instead of the straight BP to modify the network weights. The basic idea is to minimize the error between the projections of the original image and of the reconstructed image. We noticed that the system performance depends on the initial status of the network. Based on this observation, we propose a novel approach for choosing optimal values of the connection weights. The experimental results indicate that the new method can find a satisfactory solution despite that only a few projections are available.

Recent developments in Pulse-Coupled Neural Networks (PCNN) techniques provide efficiency in edge and target extraction. The detection of targets is facilitated by PCNN multi-scale image factorization. But noise is still the enemy of PCNN. An efficient new Pulse-Coupled Neural Networks technique has been proposed in combination with the wavelet theory. The new Pulse-Coupled Neural Network Wavelet (PCNNW) is based on multi-resolution decomposition for extracting the main features of the images by eliminating the noise. In addition, the wavelet coefficients provide the Pulse-Coupled Neural Network (PCNN) supplemental discrimination and lead to characteristic sets of numbers useful in identifying image factors of interest. The efficiency of the method has been tested and compared with other PCNN denoising methods.

In this paper, we investigate several fusion techniques for designing a composite classifier to improve the performance (probability of correct classification) of FLIR ATR. The motivation behind the fusion of ATR algorithms is that if each contributing technique in a fusion algorithm (composite classifier) emphasizes on learning at least some features of the targets that are not learned by other contributing techniques for making a classification decision, a fusion of ATR algorithms may improve overall probability of correct classification of the composite classifier. In this research, we propose to use four ATR algorithms for fusion. We propose to use averaged Bayes classifier, committee of experts, stacked-generalization, winner-takes-all, and ranking-based fusion techniques for designing the composite classifiers. The experimental results show an improvement of more than 5 % over the best individual performance.

Boosting has emerged as a popular combination technique to refine weak classifiers. Pioneered by Freund and Schapire, numerous variations of the AdaBoost algorithm have emerged, such as Breiman's arc-fs algorithms. The central theme of these methods is the generation of an ensemble of a weak learning algorithm using modified versions of the original training set, with emphasis placed on the more difficult instances. The validation stage then aggregates results from each element of the ensemble using some predetermined rule. In this paper the wavelet decomposition based codebook classifier proposed by Chan et al. is used as the learning algorithm. Starting with the whole training set, modifications to the training set are made at each iteration by resampling the original training data set with replacement. The weights used in the resampling are determined using different algorithms, including AdaBoost and arc-fs. Accuracy of the ensembles generated are then determined using various combination techniques such as simple voting and weighted sum.

Passive infrared imagers have long been used to detect military targets in operational scenarios. The proliferation of sensors on the battlefield has increased the need for automatic detection algorithms with low false alarm rates and high detection rates. Most infrared imagers currently operate in a single band. We are investigating the utility of dualband passive infrared sensors for target detection, and attempting to quantify the performance improvement over single band sensors. The two bands used in this research were broadband longwave and broadband midwave. The performance differences were observed using a similar set of neural-based target detectors, each of which consists of an eigenspace transformation and a simple multilayer perceptron (MLP) with different inputs. The detectors were trained with midwave-only, longwave-only, as well as signal-level and feature-level dualband inputs. Experimental results indicate significant performance improvement by the dualband inputs over single band data.

Solid waste recycling is more and more increasing according to the need to realize dismantled material recovery and to reduce overall environmental pollution. When a recycling strategy is applied sorting strategies have to be developed and implemented. Such an approach ca be considered as the second logical step of the process that is, after that the attributes (physical, chemical, morphological, morphometrical, textural, etc.) of the materials resulting from classical processing (comminution, classification, separation, etc.) are detected and numerically modeled. The resulting feature vector need to be handled by a software architecture performing the required recognition/classification procedure and defining the quality of the investigated products. From the results further feed-back or feed-forward control strategies can be applied in order to improve equipment or processing architectures performances. In this paper are analyzed and described neural network based sorting strategies applied with reference to fluff (light fraction of the materials resulting from car dismantling) recognition.