High storage capacity architecture for pattern recognition using an array of Hopfield neural networks

A new approach for the recognition of images using a two dimensional array of Hopfield neural networks is presented in this paper. In the proposed method, the N×N image is divided into sub-blocks of size M×M. Two-dimensional Hopfield neural networks of size M×M are used to learn and recognize the sub-images. All the N²/M² Hopfield modules are functioning independently and are capable of recognizing the corrupted image successfully when they work together. It is shown mathematically that the network system converges in all circumstances. The performance of the proposed technique is evaluated by applying it into various binary and gray scale images. The gray scale images are treated in a three-dimensional perspective by considering an 8-bit gray scale image as 8 independent binary images. Eight layers of binary networks are used for the recognition purpose. A Fuzzy-ART based neural network is used for the classification and labeling of the outputs in the Hopfield network. By employing the new approach, it can be seen that the storage capacity of the entire pattern recognition system would be increased to 2ⁿ where n=N²/M². Experiments conducted on different images of various sizes have shown that the proposed network structure can learn and recognize images even with 30% noise. In addition, the number of iterations required for the convergence of the network is significantly reduced and the number of synaptic weights required for the entire architecture is reduced from N⁴ to N ²M². The proposed network structure is suitable for building dedicated hardware to enable the pattern recognition in real-time due to the requirement of less number of registers to store synaptic weights and reduced number of interconnections between neurons