Neural, Parallel & Scientific Computations
A generalized cellular neural network architecture for high storage capacity pattern recognition
Second international workshop on Intelligent systems design and application
Learning using distance based training algorithm for pattern recognition
Pattern Recognition Letters
Remote sensing image classification algorithm based on hopfield neural network
ISNN'06 Proceedings of the Third international conference on Advnaces in Neural Networks - Volume Part II
| Hi-index | 0.00 |
A new approach for the recognition of images using atwo dimensional array of Hopfield neural networks ispresented in this paper. In the proposed method, the N \times Nimage is divided into sub-blocks of size M \times M. Two-dimensionalHopfield neural networks of size M \times M areused to learn and recognize the sub-images. All the N2/M2Hopfield modules are functioning independently and arecapable of recognizing the corrupted image successfullywhen they work together. It is shown mathematically thatthe network system converges in all circumstances. Theperformance of the proposed technique is evaluated byapplying it into various binary and gray scale images. Thegray scale images are treated in a three-dimensionalperspective by considering an 8-bit gray scale image as 8independent binary images. Eight layers of binarynetworks are used for the recognition purpose. A Fuzzy-ARTbased neural network is used for the classificationand labeling of the outputs in the Hopfield network. Byemploying the new approach, it can be seen that thestorage capacity of the entire pattern recognition systemwould be increased to 2n where n = N2 /M2 . Experimentsconducted on different images of various sizes haveshown that the proposed network structure can learn andrecognize images even with 30% noise. In addition, thenumber of iterations required for the convergence of thenetwork is significantly reduced and the number ofsynaptic weights required for the entire architecture isreduced from N4 to N2M2. The proposed network structureis suitable for building dedicated hardware to enable thepattern recognition in real-time due to the requirement ofless number of registers to store synaptic weights andreduced number of interconnections between neurons.