Study of neural networks to improve performance for character recognition

Three methods are proposed to improve the performance of a system based on feedforward neural networks, that is used for character recognition which are perturbed with noise up to 50%. Have been studied neural networks using backpropagation algorithm. The first element introduced in this study is to train the network in stages, using input data with noise, from 0 to 25% (steps 10,15,20,25) and then from 30% to 40% (steps 30,35,40). This approach has allowed fewer errors up to 1,2% for noise plateau 40%-50%. The second proposed method consists in introducing a vector of stability for the primary information, in stages of network training and then character recognition, with role to reduce the influence of noise on the input data. Stability vector proposed is a linear array with 5 constant parameters. The network used has 35 inputs, 16,20,25,30 or 50 nodes in the hidden layer and 26 outputs (one for each letter). The simulation and experimental results demonstrate that the proposed “vector of stability” can improve performance by decreasing the number of errors up to 1,7% for a neural network with 30 or 50 nodes in hidden layer. The third element proposed is to use reverse information (reverse the bits 0 with 1 and 1 with 0). Have been used two networks that process information. Then their outputs are analyzed and processed by a logic block. With this technique performances have been improved by decreasing the number of errors with 1%, 1.5% and 2.3% for 40%, 45% and 50% noise data.