Performing fundamental image processing operations using quantised neural networks

There has been particular interest in the hardware implementation of binary or quantised neural networks because they map well onto a hardware configuration. Among the most recent and most interesting are several optoelectronic and electronic implementations of the multilayer perceptron. In order to support this hardware development, several papers have been published which provide a discretised form for the back propagation rule in the multilayer perceptron. The authors propose an alternative approach to this problem which takes advantage of the binary structure of the neural network. In order to demonstrate the power of this algorithm, two typical image processing tasks have been addressed; edge detection in a noisy image and texture classification.