In order to improve the efficiency of the feature extraction of backpropagation (BP) learning in layered neural networks, model switching for changing the function model without altering the map is proposed. Model switching involves map preserving reduction of units by channel fusion, or addition of units by channel installation. For reducing the model size by channel fusion, two criteria for detection of the redundant channels are addressed, and the local link weight compensations for map preservation are formulated. The upper limits of the discrepancies between the maps of the switched models are derived for use as the unified criterion in selecting the switching model candidate. In the experiments, model switching is used during the BP training of a layered network model for image texture classification, to aid its inefficiency of feature extraction. The results showed that fusion and re-installation of redundant channels, weight compensations on channel fusion for map preservation, and the use of the unified criterion for model selection are all effective for improved generalization ability and quick learning. Further, the possibility of using model switching for concurrent optimization of the model and the map will be discussed.

In camera-based object recognition and classification, surface color is one of the most important characteristics. However, apparent object color may differ significantly according to the illumination and surface conditions. Such a variation can be an obstacle in utilizing color features. Geusebroek et al.'s color invariants can be a powerful tool for characterizing the object color regardless of illumination and surface conditions. In this work, we analyze the estimation process of the color invariants from RGB images, and propose a novel invariant feature of color based on the elementary invariants to meet the circular continuity residing in the mapping between colors and their invariants. Experiments show that the use of the proposed invariant in combination with luminance, contributes to improve the retrieval performances of partial object image matching under varying illumination conditions.

Particle Swarm Optimization (PSO) is a search method which utilizes a set of agents that move through the search space to find the global minimum of an objective function. The trajectory of each particle is determined by a simple rule incorporating the current particle velocity and exploration histories of the particle and its neighbors. Since its introduction by Kennedy and Eberhart in 1995, PSO has attracted many researchers due to its search efficiency even for a high dimensional objective function with multiple local optima. The dynamics of PSO search has been investigated and numerous variants for improvements have been proposed. This paper reviews the progress of PSO research so far, and the recent achievements for application to large-scale optimization problems.

We propose an algorithm for image decomposition based on Hadamard functions, realized by answer-in-weights neural network, which has simple architecture and is explored with steepest decent method. This scheme saves memory consumption and it converges fast. Simulations with least mean square (LMS) and absolute mean (AM) errors on a 128128 image converge within 30 training epochs.

Neural network pruning is a technique to obtain a fully functional subset of a redundant network for the efficiency of computation. A new method to prune a redundant three-layered neural network by means of neural element fusion" is introduced. In contrast to conventional pruning techniques that remove unimportant portions of the network, our method fuses a pair of hidden layer units so that features accumulated in both units are preserved as possible. The pair of hidden layer units to be fused is chosen by evaluating a firing similarity. This similarity measure also informs when the pruning should be stopped. The fusing method was compared with well known unit removing" methods on computer simulations. The results show that our fusing method considerably reduces the error increase due to the pruning, even in subminimal networks where conventional methods are ineffective. This enables to cut down the total cost of computation to reach the minmal network configuration.

A new neural network for locating a source by integrating data from a number of sensors is considered. The network gives a solution for inverse problems using a back-propagation algorithm with the architecture to get the solution in the inter-layer weights in a coded form Three different physical quantities are applied to the network, since the scheme has three independent ports; an input port, a tutorial port and an answer port. Our architecture is useful to estimate z" in the problem whose structure is y=f(x,z) where y is the observed data, x is the sensor position and z is the source location. The network integrates the information obtained from a number of sensors and estimates the location of the source. We apply the network to two problems of location estimation: the localization of the active nerves from their evoked potential waveforms and the localization of objects from their echoes using an active sonar system.

We propose the cube-based perceptual encryption (C-PE), which consists of cube scrambling, cube rotation, cube negative/positive transformation, and cube color component shuffling, and describe its application to the encryption-then-compression (ETC) system of Motion JPEG (MJPEG). Especially, cube rotation replaces the blocks in the original frames with ones in not only the other frames but also the depth-wise cube sides (spatiotemporal sides) unlike conventional block-based perceptual encryption (B-PE). Since it makes intra-block observation more difficult and prevents unauthorized decryption from only a single frame, it is more robust than B-PE against attack methods without any decryption key. However, because the encrypted frames including the blocks from the spatiotemporal sides affect the MJPEG compression performance slightly, we also devise a version of C-PE with no spatiotemporal sides (NSS-C-PE) that hardly affects compression performance. C-PE makes the encrypted video sequence robust against the only single frame-based algorithmic brute force (ABF) attack with only 21 cubes. The experimental results show the compression efficiency and encryption robustness of the C-PE/NSS-C-PE-based ETC system. C-PE-based ETC system shows mixed results depending on videos, whereas NSS-C-PE-based ETC system shows that the BD-PSNR can be suppressed to about -0.03dB not depending on videos.

An automatic defect classification system (ADC) for use in visual inspection of semiconductor wafers is introduced. The methods of extracting the defect features based on the human experts' knowledge, with their correlations with the defect classes are elucidated. As for the classifier, Hyperellipsoid Clustering Network (HCN) which is a layered network model employing second order discrimination borders in the feature space, is introduced. In the experiments using a collection of defect images, the HCNs are compared with the conventional multilayer perceptron networks. There, it is shown that the HCN's adaptive hyperellipsoidal discrimination borders are more suited for the problem. Also, the cluster encapsulation by the hyperellipsoidal border enables to determine rejection classes, which is also desirable when the system will be in actual use. The HCN with rejection achieves, an overall classification rate of 75% with an error rate of 18%, which can be considered equivalent to those of the human experts.

A novel neural network architecture for image texture classification is introduced. The proposed model (Kernel Modifying Neural Network: KM Net) which incorporates the convolution filter kernel and the classifier in one, enables an automated texture feature extraction in multichannel texture classification through the modification of the kernel and the connection weights by the backpropagation-based training rule. The first layer units working as the convolution kernels are constrained to be an array of Gabor filters, which achieves a most efficient texture feature localization. The following layers work as a classifier of the extracted texture feature vectors. The capability of the KM Net and its training rule is verified using a basic problem on a synthetic texture image. In addition, the possibilities of applying the KM Net to natural texture classification and biological tissue classification using an ultrasonic echo image have been tried.