Convolutional neural network algorithm design implementation method based on heterogeneous calculation

Abstract

The invention belongs to the technical field of heterogeneous calculation and image identification, and particularly relates to a convolutional neural network algorithm design implementation method based on heterogeneous calculation. According to the method, an implemented hardware platform is a Xilinx ZYNQ-7020 programmable SoC (system on a chip), an FPGA (field programmable gate array) and an ARM (advanced RISC machine) processor are arranged in the hardware platform, an implemented software platform is an SDSoC, and high-level synthesis and software definition connecting frame are combinedtogether, so that a HLS (high-level synthesis) result can be seamlessly connected to a software application. According to the method, a network model and a training network model are designed on a PC(personal computer), a network model parameter is extracted on the PC, software and hardware code partition is rapidly performed on a convolutional neural network algorithm on the SDSoC, inputted dataimage preprocessing, a pooling layer and a classification algorithm are implemented on an ARM terminal, convolution operation with maximum calculated amount is mapped to the FPGA and implemented, andperformance and area required by a system are met. According to the method, a convolutional neural network algorithm is rapidly implemented by the aid of a heterogeneous platform, the efficiency of the algorithm is greatly improved, and power consumption is greatly reduced when accuracy of the convolutional algorithm is ensured.

Description

technical field [0001] The invention belongs to the technical fields of heterogeneous computing and image recognition, in particular to a method for designing and implementing a convolutional neural network algorithm based on heterogeneous computing. Background technique [0002] With the improvement of computing power and the development of scientific computing, the application of image recognition technology is becoming more and more extensive. When the traditional image recognition technology is faced with a large number of categories and a complex environment, the recognition effect is often unsatisfactory. The deep learning algorithm model based on the convolutional neural network has become a research hotspot in the field of artificial intelligence. Compared with the traditional image recognition method, the deep learning system does not need to perform complex image preprocessing process, and the network can learn the inductive features by itself without the need for ...

AI Technical Summary

Problems solved by technology

[0010] The purpose of the method of the present invention is to find a design scheme for quickly exploring the convolutional neural network model, so as to solve the problem that the current convolutional neural network algorithm is difficult to develop on hardware, the cycle is long, and the convolutional neural network algorithm function is realized based on GPU and CPU. The problem of high consumption and low efficiency

Images