NONLINEAR ALIGNMENT OF NEURAL NET OUTPUTS FOR PARTIAL SHAPE CLASSIFICATION

A neural network approach to the partial shape classification problem is derived. Although neural networks are generally robust static pattern classifiers, they are not effective in classifying patterns with inherent temporal variations. In order to compensate for temporal variations resulting from random partial occlusion, a multi-neural network system which includes a dynamic alignment procedure at the neural net outputs is proposed. In formulating the dynamic alignment stage, a similarity measure between an input and the neural net outputs is defined. Combining the robustness of neural networks with the non-linear alignment capability of dynamic alignment results in a classifier which can tolerate high degrees of random noise and random occlusion in shapes.