Single-layer edge detector with competitive unsupervised learning

A four-neuron single-layer feedforward network designed to extract average pixel intensity, edge strength, and edge orientation in a receptive field region of an image is proposed. It uses a modified Hebb rule for unsupervised training coupled with lateral inhibition among the neurons to induce natural competition. Training involves adjustments of synaptic weights linking each neuron to the receptive field, defined as a 16×16 window, centered at any randomly chosen pixel of a 256×256 pixel training image. As the primary objective is edge detection, training is performed with images rich in edges of varying strengths and orientations. Experiments show that the weights linking each neuron organize themselves into spatially separated cells with minimum overlap. The neurons at the end of the training phase have average and gradient sensing properties. Computer simulations for the four-neuron system observing a 16×16 pixel field show that four positive-valued outputs completely define the average intensity, edge orientation, and edge strength in the field