Handwritten Digit Recognition Using Neural Networks Based On Cuda Programming

Character recognition is an important research branch in the field of pattern recognition. It divides into online character recognition and optical character recognition. Handwritten numeral recognition has always been a difficult problem in the OCR. And it is also a realistic and challenging topic. In recent years, the neural network is widely used in handwritten numeral recognition algorithm, but the implementation of neural network involves a large number of matrix and vector calculations. In front of huge data, traditional programming methods can not meet the real-time nature. INVIDIA proposed CUDA in2007. The performance of GPU increased because of the development of GPU’s programmable features. GPU computing has become a new hotspot.In order to improve the efficiency of handwritten numeral recognition, an artificial neural network based on the CUDA programming was used to accelerate the speed of handwritten numeral recognition. The method is based on the parallel nature of neural network structure (neurons in the same layer are independent of each other) and the parallel nature of matrix operations in the process of neural network training. The neural network learning process was parallelled by using CUDA C language, which can perform calculations on GPU. It greatly improves the speed of handwritten numeral recognition, and realizes real-time identification. It provides a basic approach that large-scale neural network can also compute on GPU, and pave the way for further research on GPU computing.In this paper, firstly, the image processing method of handwritten digital picture was described, including image normalization, image binarization, image smoothing, thinning and feature extraction. Then the basic knowledge of the BP neural network was described. The simulation results of handwritten numeral recognition serial algorithm based on BP neural network is given. Finally, simulation of handwritten numeral recognition parallel algorithm based on BP neural network was completed. The results from comparing the performance of parallel comparation simulation and serial comparation simulation tell us that GPU computing really improved the efficiency of neural network training.