Recognition of object orientation and shape by a rotation spreading associative neural network

Paying attention to the parietal cortex, we previously developed a rotation spreading associative neural network (RSAN net) that simultaneously recognized the orientation and shape of a binary character and a grayscale human face. The present study investigated the recognition performance of the RSAN net when the numbers of learning directions and spreading directions were changed variously, and when the input object was shifted in two-D space. Based on the data, we then developed a position invariant RSAN net in order to allow for the parallel shifting of an object. The best recognition performance of the RSAN net was obtained when the number of learning directions was equal to the number of spreading directions, taking into account the recognition characteristics and learning and recognition time. However, the RSAN net was fragile to the parallel shift of the object; the shift should be within 10 pixels in maximum. On the other hand, the position invariant RSAN net achieved position and shape invariant orientation recognition (within -60/spl deg/ to 60/spl deg/) and position and orientation invariant shape recognition of human faces (at all angles), simultaneously.