基于轻量级神经网络的人群计数模型设计

Design of Crowd Counting Model Based on Lightweight Neural Network

针对传统的卷积神经网络应用在人群计数过程中的参数众多、计算消耗大,难以在轻量级平台上实现的问题,提出一种基于轻量级神经网络的人群计数模型。模型以人群的特征提取为导向,对VGG-16网络重新部署。利用GPU完成训练,在容器化开发环境下,利用深度学习的方法进行压缩量化编码,生成轻量级神经网络,提高资源利用效率。将轻量级网络模型部署到FPGA上,完成软硬件协同推断。在Mall Dataset数据集支持下进行系统验证,实验结果表明,该系统轻量化后的均方误差可达到18.4,能效比由在PC上的0.35提高到在FPGA上的1.13,实现了轻量级神经网络的准确性及低功耗性。

A traditional convolutional neural network is difficult to be implemented on a lightweight platform due to the large number of parameters and high computational consumption in the process of crowd counting.To solve this problem,a crowd counting model based on a lightweight neural network is proposed.The VGG-16 network is redeployed based on population feature extraction.GPU is used to complete the training,and in the containerized development environment,deep learning is used to compress and quantify the coding to generate a lightweight neural network to improve the resource utilization efficiency.The lightweight network model is deployed on FPGA to complete the hardware and software collaborative inference.System verification is carried out under the support of the Mall Dataset.The experimental results prove that the mean square error quantization of the system can be up to 18.4,and the energy efficiency ratio is improved from 0.35 on PC to 1.13 on FPGA,realizing the accuracy and low power consumption of the lightweight neural network.