基于嵌入式平台的卷积神经网络压缩加速方法  

Compression Acceleration Method of Convolution Neural Networks Based on Embedded Platform

作　　者：贺彦钧 张旭博[1] HE Yanjun;ZHANG Xubo(No.30 Institute of CETC,Chengdu Sichuan 610041,China)

机构地区：[1]中国电子科技集团公司第三十研究所,四川成都610041

出　　处：《通信技术》2022年第12期1636-1641,共6页Communications Technology

摘　　要：针对一般无人车或无人机平台算力较低,无法运行较大的深度神经网络目标检测模型,或者即使能运行也无法达到实时目标检测的问题,提出了基于特定嵌入式平台的轻量级卷积神经网络压缩加速方法,在结构中引入attention机制,采用分组卷积与快速卷积结构使模型推理速度加快,并通过知识蒸馏学习当前SOTA目标检测模型Fast的目标检测能力,最后通过后统计量化方法将推理模型进一步压缩提速,让模型在保持大型目标检测网络检测精度的同时,在嵌入式平台上也达到高精度实时运行的能力。在Nano无人车平台上,使用PASCAL VOC、ImageNet数据集对压缩后的目标检测模型进行实验验证。结果表明,模型参数量减少40%,平均精度均值(mean Average Precision,mAP)仅损失0.7%,每秒帧数(Frame Per Second,FPS)提升45%,并可在无人车上实时运行。General unmanned vehicle or UAV platforms have low computing power and cannot run larger deep neural network target detection models, or even if they can, they cannot achieve real-time target detection. To address this problem, a lightweight compressive acceleration method for convolution neural networks based on a specific embedded platform is proposed. The method introduces attention mechanism in the structure, uses packet convolution and fast convolution structure to speed up the reasoning speed of the model. The target detection ability of the current SOTA target detection model Fast is studied by knowledge distillation. Finally, the reasoning model is further compressed and speeded up by a post-statistical quantification method, so that the model can maintain the detection accuracy of large target detection network and make the model run on unmanned vehicle or UAV platform with high precision. In this paper, the compressed target detection model is verified on the Nano unmanned vehicle platform using PASCAL VOC and ImageNet datasets. Experimental results indicate that the number of parameters of the model is reduced by 40%, the loss of mAP(mean Average Precision) is only 0.7%, the FPS(Frame Per Second) is increased by 45%, and it can run in unmanned vehicle in real time.

关 键 词：模型加速 知识蒸馏 目标检测 attention机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]