Application Research Of Convolutional Neural Network Based On Heterogeneous Computing Systems

With the rapid development of computer vision,deep learning has achieved great success in the fields of image,voice,text and so on.Convolution neural network,which combines deep learning technology,artificial neural network and image local correlation,can effectively extract image features,especially for computer vision tasks,and has achieved remarkable achievements in image classification,object detection,target tracking and other tasks.The convolution neural network algorithm is composed of a large number of independent multiplication and addition operation,and the computer vision task needs high performance computing processing ability,and realizes the parallel operation of the convolution neural network algorithm,which is of great significance to improve the performance of the algorithm.FPGA has strong parallel computing power.The heterogeneous computing system represented by FPGA is a good parallel computing scheme.As a cross platform development language,OpenCL provides a new and convenient way of developing parallel computing based on heterogeneous computing systems.This development mode can give full play to the parallel computing characteristics of convolution neural network,and achieve high-performance computation of convolution neural network.Based on the classical LeNet-5 model of the convolution neural network,a handwritten digital recognition system is designed,and the system parameters are trained and the sample is tested by using the MNIST data set.After analyzing the computational parallelism of the convolution neural network algorithm,the parallel acceleration scheme of the handwritten digital recognition system is designed in combination with the OpenCL technical standard and the programming method.Combining the hardware characteristics of heterogeneous computing system and the OpenCL optimization method,from three aspects of data processing optimization,storage access optimization and OpenCL library optimization,a variety of optimization schemes are designed for the handwritten digital recognition system acceleration scheme,and the performance of the convolution neural network algorithm is further optimized.In this paper,the comparison experiment of handwritten digital recognition system is carried out on X86 platform and heterogeneous computing system,and the basic scheme,acceleration scheme and optimization scheme of the system are tested respectively.Experimental data show that,in the completion of the same handwritten digital image recognition task,the parallel acceleration scheme and optimization scheme based on the heterogeneous computing system,compared with the basic scheme of the serial programming on the X86 platform,the acceleration scheme has a great performance improvement,and the optimization scheme has a larger performance increase space than the acceleration scheme.The handwritten numeral recognition system based on heterogeneous computing system design and implementation can accomplish the recognition task more efficiently.Through the study of the application of the convolution neural network algorithm for handwritten digital recognition,it is known that the heterogeneous computing system has a great advantage in the field of high performance computing.The convolution neural network will also use heterogeneous computing to further improve the computing performance and broaden the application field.