Recognition of unconstrained handwritten numerals based on dual cooperative neural network
Sukhan Lee, Yeongwoo Choi · 1994
A new neural network architecture called Dual Cooperative Neural Network (DCN) is presented in this dissertation for the recognition of totally unconstrained handwritten numeral patterns with an expectation of improved accuracy and reduced recognition time. DCN implements within its structure observations of human logical understanding and learning of numeral patterns and invariance properties which are modeled from biological visual system. The resulting structure of DCN consists of two cooperative networks: a Cartesian Network (CN) and a Log-Polar Network (LPN). The CN uses inputs represented in Cartesian coordinates, and the LPN uses the same inputs represented after the log-polar mapping. In log-polar transformed representation, rotation or scale of inputs in Cartesian coordinates appears as horizontal or vertical shifts, which can be easily detected by using nearby horizontal or vertical feature detecting cells in log-polar feature maps. Also both data representations have distinctiveness in their shapes for recognition. Each network is also hierarchically configured with three layers: a local feature map layer, a maximum selection layer, and a decision layer with back-propagation networks. DCN achieves robustness to positional shift, rotation, and scale by defining areas of feasible feature locations in both feature maps. Distortion is handled by multiple and blurred representative shapes generated by self-organization of feature maps, and is also handled by their combinations. The experimental results indicate that DCN is robust to various forms of local and global deformations in real data by achieving a 97.33% recognition rate without rejection and 94.74% with rejection for a 1% error with Zip code test numerals.