An efficient second-order neural network architecture for orientation invariant character recognition
Russell W. Duren · 1991
This dissertation presents a new second-order orientation invariant neural network architecture. The proposed architecture has two key features. The first is the use of polar sampling in combination with translation invariant second-order neurons to achieve rotation invariance without incurring the penalty of excessive network size encountered by third-order methods. The second key feature is the use of sparse connections between the input layer and the first hidden layer. It is shown that the sparsely connected architecture results in further reduction in the network size while improving the recognition rate and increases the network's resistance to problems caused by neuron saturation. These improvements are obtained without the use of subsampling thereby avoiding the resulting loss of recognition accuracy. An additional feature of the proposed architecture is that the range of rotational invariance can be tailored to fit specific applications resulting in increased recognition rates. The proposed architecture is compared to existing techniques using both theoretical and empirical methods. The theoretical comparison includes presentation of three new theorems that compare the discrimination ability of second-order neurons to translation invariant transform based methods. The empirical comparison consists of character recognition experiments using block characters and a large database of handwritten digits. Comparisons are made to methods based on geometric moments, moment invariants, Zernike moments, and conventional first-order neural networks. The experiments show that the performance of the proposed architecture equals or exceeds that of all these techniques. Additionally the proposed architecture is shown to provide computational advantages over the techniques with comparable performance.