Backpropagation algorithm for multiresolution image classification

This paper proposes a variation of the standard backpropagation (BP) algorithm that is particularly suitable for training neural networks utilized in multiresolution image classification. The aim is to solve the problem of determining which image resolutions to utilize as network inputs. The approach is to have a mix of image resolutions as network inputs and then apply the proposed BP variant. This variant not only solves the image classification problem, but also removes redundant multiresolution network inputs aiming at attaining only those necessary for classification. Removal of redundant inputs is based on that during the backward phase of BP training the network input units whose inputs are extracted from shared image space inhibit one another. Experimental results that demonstrate the viability of the proposed BP variant are presented