摘要
针对当前基于卷积神经网络的目标检测算法在小型图像处理计算平台中兼容性较差、计算能力低下以及网络训练过程中占用内存过大的问题,提出了一种轻量化卷积神经网络(CNN)YOLO-Slim,并利用YOLOv3验证可行性。首先,通过网络基础构架的改变以及将标准卷积替换为深度可分离卷积实现了网络参数与计算量的大幅度降低;其次,依据网络层对平均精度均值(mAP)的影响程度剪枝网络层,实现网络的层间剪枝;然后,使用中位数的通道剪枝策略实现对网络的层内剪枝,最终,完成轻量化网络的设计。实验结果表明,在VOC2007测试数据集上所设计的YOLO-Slim较原始YOLOv3在模型大小方面减小了90%;mAP为76.42%,识别速度为16 ms。能够为微型图像计算平台提供快速精确的目标识别能力。
Aiming at the problems of poor compatibility,low computational power and excessive memory occupied in the training process of current convolutional neural network-based target detection algorithms in small image processing computing platform,a ligntweighted convolutional neural network named YOLO-Slim was proposed and its feasibility was verified by YOLOv3.Firstly,the network parameters and computational complexity were greatly reduced by changing the network infrastructure and replacing the standard convolution with the deep separable convolution.Secondly,according to the impact of network layer on mAP(mean Average Precison),the network layer was pruned to achieve inter-layer pruning of network;then the median channel pruning strategy was used to achieve intra-layer pruning of network;finally,the lightweight network design was completed.The experimental results show that the YOLO-Slim designed on VOC2007 test data set is 90%smaller than the original YOLOv3 in model size,76.42%in mAP and 16 ms in recognition speed.It can provide fast and accurate target recognition capability for small image computing platform.
作者
邵伟平
王兴
曹昭睿
白帆
SHAO Weiping;WANG Xing;CAO Zhaorui;BAI Fan(School of Mechanical Engineering,Shenyang Ligong University,Shenyang Liaoning 110159,China;School of Equipment Engineering,Shenyang Ligong University,Shenyang Liaoning 110159,China)
出处
《计算机应用》
CSCD
北大核心
2020年第S01期8-13,共6页
journal of Computer Applications
基金
装备预研兵器工业联合基金资助项目(614B012841)。
作者简介
邵伟平(1968-),女,重庆合阳人,教授,博士,主要研究方向:机械理论及设计;通信作者:王兴(1995-),男,甘肃天水人,硕士研究生,CCF会员,主要研究方向:目标探测与识别,电子邮箱:475770828@qq.com;曹昭睿(1993-),男,辽宁沈阳人,博士研究生,主要研究方向:智能武器识别探测与制导;白帆(1982-),男,辽宁沈阳人,副教授,博士,主要研究方向:目标探测与识别。
