基于改进YOLOv4算法的轻量化网络设计与实现

doi:10.19678/j.issn.1000-3428.0060948

计算机工程 ›› 2022, Vol. 48 ›› Issue (3): 181-188. doi: 10.19678/j.issn.1000-3428.0060948

基于改进YOLOv4算法的轻量化网络设计与实现

孔维刚¹, 李文婧², 王秋艳¹, 曹鹏程³, 宋庆增¹

1. 天津工业大学计算机科学与技术学院, 天津 300387;
2. 天津工业大学电气工程与自动化学院, 天津 300387;
3. 中国电子科技集团公司信息科学研究院, 北京 100086

收稿日期:2021-02-26 修回日期:2021-04-15 发布日期:2021-04-27
作者简介:孔维刚(1997-),男,硕士研究生,主研方向为FPGA开发、嵌入式系统开发、深度学习;李文婧,硕士研究生;王秋艳,副教授、博士;曹鹏程,博士;宋庆增(通信作者),副教授、博士。
基金资助:
国家自然科学基金（61802281，61702366）；天津市自然科学基金（18JCQNJC70300，19JCYBJC15800）；天津市教委科研计划项目（2018KJ215，2020KJ112，KYQD1817）。

Design and Implementation of Lightweight Network Based on Improved YOLOv4 Algorithm

KONG Weigang¹, LI Wenjing², WANG Qiuyan¹, CAO Pengcheng³, SONG Qingzeng¹

1. School of Computer Science and Technology, Tiangong University, Tianjin 300387, China;
2. School of Electrical Engineering and Automation, Tiangong University, Tianjin 300387, China;
3. Information Science Academy of China Electronics Technology Group Corporation, Beijing 100086, China

Received:2021-02-26 Revised:2021-04-15 Published:2021-04-27

摘要/Abstract

摘要： 在嵌入式设备上进行目标检测时易受能耗和功耗等限制，使得传统目标检测算法效果不佳。为此，对YOLOv4算法进行优化，设计YOLOv4-Mini网络结构，将其特征提取网络由CSPDarkNet53改为MobileNetv3-large并进行INT8量化处理，其中网络结构利用PW和DW卷积操作代替传统卷积操作以大幅减少计算量。采用SE模块为通道施加注意力机制，激活函数层运用h-swish非线性激活函数，在保证精度的情况下降低网络计算量。同时，通过量化感知训练将权重转为INT8类型，以实现模型轻量化，进一步降低网络参数量和计算量，从而在嵌入式设备上完成无人机数据集的目标检测任务。在NVIDIA Jetson Xavier NX设备上进行测试，结果显示，YOLOv4-MobileNetv3网络的mAP为34.3%，FPS为30，YOLOv4-Mini网络的mAP为32.5%，FPS为73，表明YOLOv4-Mini网络能够在低功耗、低能耗的嵌入式设备上完成目标实时检测任务。

关键词: 目标检测, 模型压缩, 嵌入式设备, 轻量化神经网络, 模型量化, Jetson Xavier NX设备

Abstract: Target detection using embedded devices is limited by energy and power consumption, deteriorating the performance of traditional target detection algorithms.Therefore, to address this issue, the YOLOv4 algorithm is optimized;the YOLOv4-Mini network structure is designed;the feature extraction network is changed from CSPDarkNet53 to MobileNetv3-large;and INT8 quantization processing is carried out.The network structure uses PW and DW convolution operations to replace the traditional convolution operation to greatly reduce the amount of calculation.The SE module is used to apply attention mechanism to the channel, and h-swish nonlinear activation function is used in the activation function layer to reduce the amount of network calculation while ensuring accuracy. Concurrently, the weight is transformed into INT8 type through quantitative perception training to realize the lightweight of the model and further reduce the amount of network parameters and computation, in addition to completing the target detection task of UAV data set on embedded devices.The test results on NVIDIA Jetson Xavier NX show that the mAP of YOLOv4-MobileNetv3 network is 34.3%, the FPS is 30, the mAP of YOLOv4-Mini network is 32.5%, and the FPS is 73 indicating that the YOLOv4-Mini network can complete the target real-time detection task on the embedded device with low power consumption and low energy consumption.

Key words: target detection, model compression, embedded device, lightweight neural network, model quantification, Jetson Xavier NX equipment

中图分类号:

TP391

孔维刚, 李文婧, 王秋艳, 曹鹏程, 宋庆增. 基于改进YOLOv4算法的轻量化网络设计与实现[J]. 计算机工程, 2022, 48(3): 181-188.

KONG Weigang, LI Wenjing, WANG Qiuyan, CAO Pengcheng, SONG Qingzeng. Design and Implementation of Lightweight Network Based on Improved YOLOv4 Algorithm[J]. Computer Engineering, 2022, 48(3): 181-188.

http://www.ecice06.com/CN/Y2022/V48/I3/181

图/表 11

20220331202347

20220331202350

20220331202353

20220331202356

20220331202359

20220331202402

20220331202405

20220331202408

20220331202411

20220331202416

20220331202419

参考文献

[1] GIRSHICK R, DONAHUE J, DARRELL T, et al.Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of IEEE Conference on Computer Vision & Pattern Recognition.Washington D.C., USA:IEEE Computer Society, 2014:580-587.
[2] GIRSHICK R.Fast R-CNN[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2015:1440-1448.
[3] REN S, HE K, GIRSHICK R, et al.Faster R-CNN:towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6):1137-1149.
[4] LIU W, ANGUELOV D, ERHAN D, et al.SSD:single shot multibox detector[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2016:21-37.
[5] REDMON J, DIVVALA S, GIRSHICK R, et al.You Only Look Once:unified, real time objecct detection[C]//Proceedings of European Conference on Computer Vision and Pattern Recognition.Berlin, Germany:Springer, 2017:6517-6525.
[6] REDMON J, FARHADI A.YOLO9000:better, faster, stronger[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2017:6517-6525.
[7] REDMON J, FARHADI A.YOLOv3:an incremental improvement[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2018:1804-1823.
[8] BOCHKOVSKIY A, WANG C Y, LIAO H Y M.YOLOv4:optimal speed and accuracy of object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:102-123.
[9] HOWARD A, SANDLER M, CHU G, et al.Searching for MobileNetv3[EB/OL].[2021-01-05].https://arxiv.org/pdf/1905.02244v3.pdf.
[10] ZHANG Z, HE T, ZHANG H, et al.Bag of freebies for training object detection neural networks[EB/OL].(2019-04-12)[2021-01-05].https://arxiv.org/abs/1902.04103.
[11] JIE H, LI S, GANG S.Squeeze-and-excitation networks[EB/OL].[2021-01-05].https://arxiv.org/pdf/1709.01507.pdf.
[12] HE K, ZHANG X, REN S, et al.Spatial pyramid pooling in deep convolutional networks for visual recognition[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9):1904-1916.
[13] HOWARD A G, ZHU M, CHEN B, et al.MobileNets:efficient convolutional neural networks for mobile vision applications[EB/OL].[2021-01-05].https://arxiv.org/pdf/1704.04861.pdf.
[14] SANDLER M, HOWARD A, ZHU M, et al.MobileNetv2:inverted residuals and linear bottlenecks[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2018:4510-4520.
[15] JACOB B, KLIGYS S, CHEN B, et al.Quantization and training of neural networks for efficient integer-arithmetic-only inference[EB/OL].[2021-01-05].https://arxiv.org/pdf/1712.05877.pdf.
[16] CHEN Y H, EMER J, SZE V.Eyeriss:a spatial architecture for energy-efficient dataflow for convolutional neural networks[EB/OL].[2021-01-05].https://www.rle.mit.edu/eems/wp-content/uploads/2016/04/eyeriss_isca_2016.pdf.
[17] JOUPPI N P, YOUNG C, PATIL N, et al.In-datacenter performance analysis of a tensor processing unit[C]//Proceedings of 2017 ACM/IEEE Annual International Symposium on Computer Architecture.Washington D.C., USA:IEEE Press, 2017:1-12.
[18] JOUPPI N, YOUNG C, PATIL N, et al.Motivation for and evaluation of the first tensor processing unit[J].IEEE Micro, 2018, 38(3):10-19.
[19] MISRA D.Mish:a self regularized non-monotonic neural activation function[EB/OL].[2021-01-05].https://arxiv.org/pdf/1908.08681.pdf.
[20] SERRA T, KUMAR A, RAMALINGAM S.Scaling up exact neural network compression by ReLU stability[EB/OL].[2021-01-05].https://arxiv.org/pdf/2102.07804.pdf.
[21] ZHANG S H, LI X L, ZONG M, et al.Learning k for kNN classification[J].ACM Transactions on Intelligent Systems and Technology, 2017, 8(8):1-19.

选择文件类型/文献管理软件名称

选择包含的内容

基于改进YOLOv4算法的轻量化网络设计与实现

Design and Implementation of Lightweight Network Based on Improved YOLOv4 Algorithm

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献

相关文章 15

编辑推荐

Metrics

本文评价

[1]	宋志娜, 李莎, 杨建明, 徐川. 基于特征与区域定位增强的遥感舰船目标检测[J]. 计算机工程, 2023, 49(8): 257-264.
[2]	刘俊豪, 王美林, 谢兴, 宋烨兴, 许莉花. 基于改进YOLOv5的皮革瑕疵检测算法[J]. 计算机工程, 2023, 49(8): 240-249.
[3]	李强龙, 周新文, 位梦恩, 甘阳洲. 基于条形池化和注意力机制的街道场景红外目标检测算法[J]. 计算机工程, 2023, 49(8): 310-320.
[4]	徐春波, 闫娟, 杨慧斌, 王博, 吴晗. 基于目标检测和语义分割的视觉SLAM算法[J]. 计算机工程, 2023, 49(8): 199-206, 214.
[5]	闫兴亚, 匡娅茜, 白光睿, 李月. 基于深度学习的学生课堂行为识别方法[J]. 计算机工程, 2023, 49(7): 251-258.
[6]	李军侠, 王星驰, 殷梓, 石德硕. 边缘深度挖掘的弱监督显著性目标检测[J]. 计算机工程, 2023, 49(7): 169-178.
[7]	吴珊, 周凤. 基于改进SSD算法的小目标检测[J]. 计算机工程, 2023, 49(7): 179-188.
[8]	齐咏生, 杜晓旭, 朱俊峰, 高胜利, 刘利强. 基于增强型轻量深度网络的牧区牲畜高效检测[J]. 计算机工程, 2023, 49(7): 278-287.
[9]	聂志勇, 阴宇薇, 汤佳欣, 涂志刚. 一种基于边界框关键点距离的框回归算法[J]. 计算机工程, 2023, 49(7): 65-75.
[10]	谌雨章, 黄逸姿, 张钧涵. 基于多速率空洞卷积的多尺度水下小目标检测[J]. 计算机工程, 2023, 49(6): 257-264.
[11]	李宜亭, 屈丹, 杨绪魁, 张昊, 沈小龙. 基于分解门控注意力单元的高效Conformer模型[J]. 计算机工程, 2023, 49(5): 73-80.
[12]	罗华峰, 沈奕菲, 阮黎翔, 杜奇伟, 郑翔, 陈智麒, 张胜. 边缘环境下面向实时目标检测的帧卸载调度算法[J]. 计算机工程, 2023, 49(5): 295-301,309.
[13]	郭奕裕, 周箩鱼. 安全帽佩戴检测网络模型的轻量化设计[J]. 计算机工程, 2023, 49(4): 312-320.
[14]	王璐璐, 陈东方, 王晓峰. 一种基于锚框质量分布的动态标签分配策略[J]. 计算机工程, 2023, 49(4): 85-91,100.
[15]	宋鹏鹏, 龚声蓉, 钟珊, 周立凡, 凤黄浩. 基于双注意力擦除和注意力信息聚合的弱监督目标检测[J]. 计算机工程, 2023, 49(3): 113-120,127.

模态框（Modal）标题

选择文件类型/文献管理软件名称

选择包含的内容

基于改进YOLOv4算法的轻量化网络设计与实现

Design and Implementation of Lightweight Network Based on Improved YOLOv4 Algorithm

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献

相关文章 15

编辑推荐

Metrics

本文评价