A biologically motivated model for rotation, shift, scaling and distortion invariant pattern recognition

Summary form only given. A biologically motivated multilayer neural network for 2D visual pattern recognition with rotation, shift, scaling, and distortion invariance is proposed. Based on the hierarchical module proposed by Hubel and Wiesel, the new structure has a rotation layer in the image preprocessing area, followed by a neocognitron-type structure and a modified masking fields neural network module in the associative area. The rotation layer rotates the image to different orientations on polar coordinates to solve the rotation problem of an input object. By combining with a neocognitron-type network, the structure possesses shift, scaling, and distortion invariant properties. A modified masking fields network allows competition and cooperation with different excited components to make the classification for patterns without significant distortion and scaling, and find out the orientation of an input pattern. Based on the detected orientation, a second associative module similar to the neocognitron is incorporated to further resolve significant distorted and scaled patterns. Experimental results have been obtained for target and hand-written character recognition