The Improvement Of BP Neural Network And Its Application In Handwritten Recognition

Nowadays, handwritten digit recognition technique is researched importantly. In traditional digit recognition, prototype method is only adaptive to print digit recognition; statistics decision method is difficult reflect structure feature elaborately; word structure has weak resisting other disturbance information. In this article, I use BP neural network to recognize handwritten digit samples. According to network's strong nonlinear digit treatment ability, after preprocessing and feature extraction, we only offer numerous sample digit matrix to network, and digit classification is obtained. But traditional BP neural network has many defects and can't be used directly; the selection of network parameter and decision is complex and vital. I analyze and compare the improvement based on traditional gradient descent methods and numerical optimization thoughts respectively. After that, we use fault pattern recognition experiment to research the training and recognition performance of different improved algorithms. In the method of momentum way and self-adaptive learning velocity based on gradient descent, training figure is smooth and training course is easy to be controlled, recognition result is better too, but in large network, training velocity is slow. LM-BP algorithm based on numerical optimization LM-BP algorithm has fastest training epochs and best recognition performance, but its training is difficult to control, what's more, it needs big computer inner storage.In this article, I use different hidden node selection methods and several improved methods to do handwritten digit experiment under 256M computer inner environment. As the handwritten digit samples are special, preprocessing and feature extraction are analyzed and argued elaborately. After that course, we select 600 standard handwritten digit samples and construct network. I use a kind of 13-dimension sample structure feature vector to reflect sample information. According to my programs under MATLAB 7.0, after training and recognition testing, when some formula is selected and 27 hidden nodes are used, the performance is best. The momentum and adaptive learning velocity algorithm based on gradient descent method and LM-BP algorithm based on numerical optimization algorithms has much better performance, faster training velocity and higher recognition rate than others, their recognition rates can rearch about 90%.They are compared by some table or figure formats.