Research on Edge End Food Recognition Technology based on ZYNQ

Abstract

Aiming at the problems of real-time and poor security of traditional restaurant specific food recognition technology, this paper proposes an edge food recognition method based on ZYNQ platform. In this method, a dataset containing 8166 Chinese food images is first established and the YOLOv2 food recognition model is trained. Then, the IP core of YOLOv2 accelerator is designed. Next, the IP-based design process is combined with hardware optimization. Through quantization, software and hardware task division, deployment and other steps, the trained YOLOv2 food recognition model is deployed to ZYNQ platform, and the real-time recognition of 20 kinds of Chinese food pictures is finally realized. The experimental evaluation results show that compared with the existing CNN target detection method based on FPGA platform, the proposed scheme has obvious advantages in energy efficiency ratio, and the recognition speed is about 1FPS, and the mAP classification of 20 kinds of Chinese food image reaches 0.882.