Engineers have utilised spectral information and have steadily improved its applications in imaging systems for more than a century. The course of technological developments in colour imaging has been dictated by system improvements measured by their efficacy for direct human consumption. It seems reasonable to us to try to emulate nature and boost capabilities of machine vision systems by optimising the way in which they exploit spectral information. This is a two-step process: first step involves using a few spectrally broad detectors to compress the information content of the scene and the second step constructs spectral discriminants for image segmentation based on a small number of spectrally generated features assigned to each pixel. In animals the discriminant value is attributed to the object as what is called colour. Previous papers have concentrated on the final segmentation step. Here we show a straightforward way to design application-specific spectral sensitivity functions to improve image segmentation. The resulting functions can be used for reliable recognition of objects in a hyperspectral image in real time. These functions can also be used to design task-specific specialised cameras that can outperform current hyperspectral systems in terms of sensitivity, size, power consumption, robustness, price and complexity.

Conscious behavior is hypothesized to be governed by the dynamics of the neural architecture of the brain. A general model of an artificial consciousness algorithm is presented, and applied to a one-dimensional feedback control system. A new learning algorithm for learning functional relations is presented and shown to be biologically grounded. The consciousness algorithm uses predictive simulation and evaluation to let the example system relearn new internal and external models after it is damaged.

A pulse coupled neural network (PCNN) can run mazes nondeterministically (taking all possible paths) with constant time per step. Thus, when a signal emerges, it has taken the shortest path in the shortest time.

Fourier optics has offered a powerful global information analysis and processing tool to optical domain signals and images and firmly locked its respectable position in the history of optics. The evolution of wavelet theory in the past decade, which has created significant impacts on many areas of information science and engineering, has recently been frequently mentioned in the optics community. Whether or not this locality-oriented signal analysis and processing tool can play a comparably significant role in optics remains to be answered. The present paper is intended to summarize important developments and recent progress in areas where novel wavelet processing concepts are incorporated into optics research.

When operating in the real world, the conventional fuzzy approach is to extract crisp data then fuzzify it. This paper explores the concept of measuring fuzzy set memberships directly and obtaining the underlying crisp values by defuzzification. Both speed and accuracy advantages of fuzzy over conventional, crisp metrology are noted.

One of the main problems in current (artificial) neural network engineering is the lack of design rules for neural networks, i.e., how many layers and how many neurons per layer to choose for a fully connected layered neural network with bidirectional weights. A theory is developed which optimizes the topology of the neural network to allow a maximum potential storage capacity with a minimum amount of neurons. The optimal topology and maximum number of connections for all the interconnection schemes are given

It is shown that full interconnection of an N*N array to an N*N output array with a rank-1 N/sup 4/-component matrix is quite straightforward for essentially all values of N. A number of applications and implementations are then described, showing that for fixed interconnect systems one can achieve small size, ruggedness, low cost, and speed.>

Classical optical information processing and classical neural networks can be adapted and combined to create optical neural networks which offer significant and fundamental advantages over electronic neural networks in various well-defined cases. A systematic morphology of optical neural networks is presented. Special problems they create are discussed. The state of the art of their implementation is indicated, and some supportable speculations on their future are given.>