基于Zynq平台的低功耗人脸检测加速系统

Low-power Face Detection Acceleration System Based on the Zynq Platform

摘要:

基于CPU及GPU的卷积神经网络平台存在体积大、能耗高等问题,提出了一种基于Zynq平台的卷积神经网络人脸检测加速系统。该系统采用YOLOv3-Tiny算法,并利用Wider Face人脸数据集进行训练。为提高网络效率,采用层融合技术减小网络深度,加快检测速度；同时,采用8位整数量化策略,以降低内存访问量,减少资源消耗。通过利用ZynqXC7Z035芯片上FPGA端并行计算能力,设计出可重复利用的多通道卷积计算模块,实现DSP的重复递用。实验结果显示,所设计的加速系统实现了9.5FPS的实时推理速度,检测速度是intel i7-8700CPU的7.9倍,系统功耗仅为2.65W,满足低功耗的性能需求。

Abstract:

To address the issues of large size and high power consumption in CPU- and GPU-based convolutional neural network platforms, we designed and implemented a convolutional neural network-assisted face detection acceleration system based on the Zynq platform in this study. We adopted the YOLOv3-Tiny algorithm for the proposed system and used the WIDER FACE dataset for training. To improve the network efficiency, we utilized a layer-fusion technique for reducing the network depth and accelerating detection. Moreover, we employed an 8-bit integer quantization strategy to minimize memory access and resource consumption. We designed a reusable multichannel convolution computation module by leveraging the parallel computing capability of field-programmable gate arrays (FPGAs) on the ZynqXC7Z035 chip to reuse the digital signal processor (DSP). The experimental results showed that our designed acceleration system, which could achieve a real-time inference speed of 9.5FPS, was 7.9 times faster than intel i7-8700CPU and consumed only 2.65W of power, satisfying the performance requirement of low power consumption.