PATTERN-RECOGNITION PROPERTIES OF VARIOUS FEATURE SPACES FOR HIGHER-ORDER NEURAL NETWORKS

Higher order neural networks (HONN's) are a form of preprocessing for the standard backpropagation neural networks that use geometrically motivated nonlinear combinations of scene pixels to achieve invariant pattern recognition feature spaces. In standard backpropagation, scene pixel values are presented directly to the neural network input nodes. By proper choice of HONN pixel combinations, it is possible to directly incorporate geometric invariance properties into the HONN. The HONN can be considered to be a particular type of preprocessing that explicitly creates nonlinear decision surfaces. Originally, HONN's had fully interconnected input pixels that caused a severe storage requirement. We explore alternatives that reduce the number of network weights while maintaining geometric invariant properties for recognizing patterns in real-time processing applications. This study is limited to translation and rotation invariance. We are primarily interested in examining the properties of various feature spaces for HONN's, in correlated and uncorrelated noise, such as the effect of various types of input features, feature size and number of feature pixels, and effect of scene size. We also consider the HONN training robustness in terms of target detectability. The experimental setup consists of a 15 x 20 pixel scene possibly containing a 3 x 10 target. Each trial used 500 training scenes plus 500 testing scenes. Results indicate that HONN's yield similar geometric invariant target recognition properties to classical template matching. However, the HONN's require an order of magnitude less computer processing time compared with template matching. For our simple target, HONN's with zero hidden layers yield results equivalent to HONN's with multiple layers. This reduces network training time over the multiple layer networks. Finally HONN's exhibit robust training characteristics. The HONN's that were trained at one input noise level and tested at a different level maintained similar probability of detection and probability of false alarm characteristics than did the HONN's that were trained and tested at the same input noise levels. Results indicate that HONN's could be considered for real-time target recognition applications.