The Improved RBF Neural Network And Its Application In Character Recognition

The Radial Basis Function (RBF) method is a technology of interpolation in a high-dimensional space, which is at developing stage used to the neural network (NN). A RBFNN is a kind of feed forward network which basically involves 3 layers. It is proved that its speed of convergence of is much quicker than general BP algorithm. And its structure can be fixed in the algorithm. Therefore the study and application of RBFNN is getting fast development in recent years.At present, a large number of learning algorithms have emerged. In commonly, K-means and orthogonal least square are used to train the parameters of the hidden layer, gradient-descent and recursive least square are used to train the linear weights, etc. The main problems in designing a RBFNN depend on fixing the nodes of the hidden layer, the parameters of the centers and the linear weights.This dissertation has mainly launched the following work to the above-mentioned problems.Firstly some hidden layer nodes are initialized, and train the network using competitive learning algorithm. All centers will compete when inputs a sample, the nearest one getting win and been upgraded while the losers keep its value. If this sample is far away from all the centers it will be set asa new center. Stop until convergence or reach the maximum epochs, confirm all centers and assign the samples to the corresponding centers.After clustering, the gradient-decent is used to train the weights which make the cost function minimized. Because of its light calculate, the network convergence quickly through adjusting suitable learning rate.Otherwise, to save the resource Akaike's final prediction error standard (FPE) is employed to delete the nodes that contribute little to the outputs of the network. This will balance the precision with the complexity of the network. Until the FPE no longer drops a group of final optimum weights and a rational network have been found.In this dissertation, RBFNN is applied to classify the characters. Firstly select 5400 number sample images (each half of train data and test data) and code them by a shadow mask, pick up their characters. Then train the RBFNN with the train data. Finally input the test data to the learned network to recognize, a high recognition rate of 95.2 percents gained.