FPGA-Based Heterogeneous Accelerated Image Classification System

Abstract:
This paper focuses on h eterogeneous accelerated image classification system based on FPGA. In order to ensure the system to function well in harsh environment, it is of great significance to achieve the goal of low power consumption, ensuring the detection speed and accuracy at the same time. In this paper, YOLOv2 algorithm is applied for interactive development based on FPGA-SOC. Through the reasonable allocation of resources, the total power consumption is 2.345W on the ZYNQ platform, and the classification speed can attain 3.6 FPS. Moreover, the average detection accuracy of the target is higher than 70%. It can efficiently solve the problems that the traditional high-performance computing platform consumes power highly when maintaining the speed of high frame rate detection, and the convolution neural network algorithm running on the embedded platform is extremely slow and takes up too many resources.