[1]张利红,蔡敬菊.基于轻量化 Yolov5 算法的目标检测系统[J].计算机技术与发展,2022,32(11):134-139.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 020]
　ZHANG Li-hong,CAI Jing-ju.Target Detection System Based on Lightweight Yolov5 Algorithm[J].,2022,32(11):134-139.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 020]
点击复制

基于轻量化 Yolov5 算法的目标检测系统()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年11期
页码:
134-139
栏目:
人工智能
出版日期:
2022-11-10

文章信息/Info

Title:
Target Detection System Based on Lightweight Yolov5 Algorithm
文章编号:
1673-629X(2022)11-0130-06
作者:
张利红1; 2 ; 蔡敬菊1
1. 中国科学院 光电技术研究所,四川 成都 610209;
2. 中国科学院大学,北京 100049
Author(s):
ZHANG Li-hong1; 2 ; CAI Jing-ju1
1. Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610209,China;
2. University of Chinese Academy of Sciences,Beijing 100049,China
关键词:
目标检测网络; 深度可分离卷积; 模型量化; 减枝; 硬件加速; 嵌入式部署
Keywords:
target detection network; deep separable convolution; model quantization; pruning; hardware acceleration; embedded deployment
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 11. 020
摘要:
针对现有的深度目标检测算法结构复杂、计算量过大,难以直接部署到资源有限的边缘设备进行实时检测应用的问题,以 Yolov5 算法为基础,针对 VOC 公开数据集在 GPU 上进行迭代训练, 通过使用 MobileNetv2 替换 Backbone 特征提取层中的 BottleneckCSP 结构、Conv 替换 Focus 模块达到网络轻量化,并结合稀疏训练评价特征提取层中卷积核的重要性后进行减枝的方法进一步实现模型压缩。 从模型适应平台硬件加速角度出发,根据瑞芯微 Rk3399pro 加速芯片 MAC 单元为3 的倍数,提出将网络卷积通道数剪枝后约束为 9 的倍数,并引入了非对称 8 位模型量化、CPU-GPU-NPU 多核协同工作的策略在嵌入式平台上进行 C++算法部署。 实验证明,轻量化的 Yolov5 算法在检测精确度 mAP 下降 6. 74 的情况下,大幅减少了计算参数量,离线模型部署至 Rk3399pro 嵌入式平台上理论检测速度达到 50 fps / s,相较原 Yolov5s 未优化改进的部署至平台上的速度提升近 1. 7 倍;满足降低模型参数权重后仍能实时精确检测的效果。
Abstract:
The existing deep target detection algorithm has complex structure and large calculation amount,which is difficult to directlydeploy to edge devices with limited resources for real - time detection. For this, based on the Yolov5 algorithm, iterative training isperformed on the GPU for the VOC public data set. The MobileNetv2 is used to replace the BottleneckCSP structure in the Backbonefeature extraction layer,and Conv to replace the Focus module for network lightweight. And combined with sparse training to evaluate theimportance of the convolution kernel? ?of the feature extraction layer and then pruning,the model compression is completed further. Fromthe perspective of model adaptation platform hardware acceleration,based on Rockchip’s Rk3399pro acceleration chip MAC unit being amultiple of 3,the network convolution channel number is constrained to multiples of 9 after pruning,and the asymmetric 8 - bit modelquantization and the CPU-GPU-NPU multi-core collaboration strategy is introduced to deploy C++ algorithm on embedded platform.The experiment shows that the lightweight Yolov5 algorithm greatly reduces the amount of calculation parameters when the detectionaccuracy mAp drops by 6. 74. The theoretical detection speed of offline model deployed on the Rk3399pro embedded platform reaches 50fps / s,which is nearly 1. 7 times faster than the original Yolov5s unoptimized and improved deployment to the platform. It satisfies theeffect of real-time accurate detection after reducing the weight of model parameters.

相似文献/References:

[1]侯枘辰,刘 瑜,廉 华,等.用于草坪场景理解的轻量化图像分割算法[J].计算机技术与发展,2020,30(10):59.[doi:10. 3969 / j. issn. 1673-629X. 2020. 10. 011]
　HOU Rui-chen,LIU Yu,LIAN Hua,et al.Lightweight Image Segmentation for Lawn Scene Understanding[J].,2020,30(11):59.[doi:10. 3969 / j. issn. 1673-629X. 2020. 10. 011]
[2]余 进,史燕中,王春华,等.一种轻量化目标检测算法研究[J].计算机技术与发展,2020,30(11):42.[doi:10. 3969 / j. issn. 1673-629X. 2020. 11. 008]
　YU Jin,SHI Yan-zhong,WANG Chun-hua,et al.Research of a Lightweight Object Detection Algorithm[J].,2020,30(11):42.[doi:10. 3969 / j. issn. 1673-629X. 2020. 11. 008]
[3]刘　鑫,赵慧杰,武苏雯.基于Ｕ－Ｎｅｔ的岩层图像裂缝检测算法应用[J].计算机技术与发展,2021,31(05):73.[doi:10. 3969 / j. issn. 1673-629X. 2021. 05. 013]
　,,et al.ＡｐｐｌｉｃａｔｉｏｎｏｆＵ－ＮｅｔＢａｓｅｄＲｏｃｋＦｏｒｍａｔｉｏｎＩｍａｇｅＣｒａｃｋＤｅｔｅｃｔｉｏｎＡｌｇｏｒｉｔｈｍ[J].,2021,31(11):73.[doi:10. 3969 / j. issn. 1673-629X. 2021. 05. 013]
[4]谭芳喜,肖世德,周亮君,等.基于改进 YOLOv3 算法在道路目标检测中的应用[J].计算机技术与发展,2021,31(08):118.[doi:10. 3969 / j. issn. 1673-629X. 2021. 08. 020]
　TAN Fang-xi,XIAO Shi-de,ZHOU Liang-jun,et al.Application in Road Target Detection Based on ImprovedYOLOV3 Algorithm[J].,2021,31(11):118.[doi:10. 3969 / j. issn. 1673-629X. 2021. 08. 020]
[5]徐 丽,刘星星,屈立成.基于 YOLO v4 的夜间车辆检测模型轻量化研究[J].计算机技术与发展,2022,32(03):84.[doi:10. 3969 / j. issn. 1673-629X. 2022. 03. 014]
　XU Li,LIU Xing-xing,QU Li-cheng.Research on Lightweight of Night Vehicle Detection Model Based on YOLO v4[J].,2022,32(11):84.[doi:10. 3969 / j. issn. 1673-629X. 2022. 03. 014]
[6]王烨奎,曹铁勇,王 杨,等.CAMOU-YOLO:一种迷彩伪装目标检测模型[J].计算机技术与发展,2022,32(12):29.[doi:10. 3969 / j. issn. 1673-629X. 2022. 12. 005]
　WANG Ye-kui,CAO Tie-yong,WANG Yang,et al.CAMOU-YOLO:A Camouflaged Object Detection Model[J].,2022,32(11):29.[doi:10. 3969 / j. issn. 1673-629X. 2022. 12. 005]
[7]朱万鹏,雷秀娟.使用胸片检测新型冠状肺炎的深度集成网络[J].计算机技术与发展,2023,33(02):153.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 023]
　ZHU Wan-peng,LEI Xiu-juan.Deep Integrated Network for Detection of COVID-19 Using Chest Radiographs[J].,2023,33(11):153.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 023]

更新日期/Last Update: 2022-11-10