A Prallel Algorithm Recognition Based On Artificial Neural Network With GPU-Acceleration

The character recognition is a very important research branche in pattern recognition, and handwritten digits recognition is especially significant and challenging. Resently years, the Artificial Neural Network (ANN) has been applied to this area, but with more matrix-vector calculations. Traditional parallel artificial neural network is often inefficient as it needs the parallel computers that are relatively difficult to use and may not be accessible to most researchers. With the development of GPU's performance and its programmable features, General-Purpose computation on Graphics Processing Units (GPGPU) is becoming a new research hotspot.Therefore, this paper raises a handwritten digits recognition algorithm based on ANN and GPU-acceleration. According to the separation of training and implementation of ANN and neurons in the same layer are independent of each others, this algorithm improves the generalization power of recognition network, and gets a good recognition effect with parallel acceleration computing by GPU.The algorithm adopts some methods, such as predistortion and second order, when training this recognition network, comprehensively consideratiing the change of learning rate, modifies the recogonition model using 5-layer convolutional ANN, therefore achieves a high accuracy rate. For the data calculation speedup, this paper uses GPU to compute the recognition process. By taking advantage of the parallelism of graphics hardware, this method greatly improving computing speed, reduces the time cost. This paper also provides a general way to parallelize ANN on GPU.The experimental results show that these methods can get a higher accuracy rate and a better speed performance while using low-cost hardware.