A compact representation of binary patterns for invariant recognition

In this paper, a very compact coding scheme for binary visual patterns based on a novel approach dubbed shadow codes is proposed, and the applicability of the method to invariant recognition of handwritten patterns by neural networks is investigated. In the best configuration so far, the input pattern is surrounded by a rectangular frame with orientation given by the pattern's principal axes of inertia, and then a shadow vector is obtained by projecting the pixels of the pattern into bars of the frame. After normalization, the resulting vector is fed into a rotation-invariant network, whose output is used for classification by a neural network. For a task involving the recognition of handwritten digits, experimental results with three neural network approaches, namely, self-organizing map, learning vector quantization and multilayer perceptron, show that though very compact, the proposed scheme is effective for translation, rotation, and scaling invariant recognition of simple binary patterns.