改进YOLOv3-SPP的无人机目标检测模型压缩方案

doi:10.3778/j.issn.1002-8331.2007-0230

摘要/Abstract

摘要：

无人机设备算力低下，深度模型计算量过大不适合直接部署，航拍图像目标小并且密集，模型对目标识别分类效果也不佳。为了提高深度模型航拍目标检测的精度和速度，降低计算量。对YOLOv3-SPP模型进行改进，将GIoU代替平方和用作定位损失，提高定位精度。提出了一种数据集优化和数据增强方法。再针对特定类别按照权值进行采样处理均衡化类别数量。随机组合不同场景样本组成批训练，提高模型训练效率和检测鲁棒性。再对模型进行压缩，在BN层添加缩放因子进行稀疏训练和通道剪枝的基础上，通过缩放因子衡量模型残差层重要性，修剪不重要残差，进一步减小前向推理层数和参数。实验表明，模型参数量减小了95.7%，模型大小减小95.82%，同等算力下模型推理速度提高为原来3倍。且精度和速度均高于最新YOLOv5系列轻量模型。

关键词: 目标检测, 无人机, 通道剪枝, 模型压缩, YOLOv3

Abstract:

Due to the limited memory and computing power of UAV equipment, it is very challenging inrunning deep-learning models on such devices for object detection. Dense and small sizes objects in aerial image and datasets increases the difficulty on aerial target recognition and classification. To deal with these challenges, the YOLOv3-SPP model is improved by using GIoU instead of the mean-square error, which raise the positioning accuracy if objects. A method of data enhancement is also proposed. It balances categories number by weighted sampling specific categories. During training, different scene pictures are randomly combined to form batch training, which increases detection robustness. Then, the model is compressed. On the basis of sparse training and channel pruning by adding scaling factor to BN layer, the importance of residual layer is measured by scaling factor, and unimportant residual is pruned to further reduce the number of forward reasoning layers and parameters. In experiment, model parameters are reduced by 95.7%, the module size is reduced by 95.82%, and the speed is increased by 3 times. The accuracy and speed are higher than the latest YOLOv5 series lightweight model.

Key words: object detection, UAVs, channel pruning, model compression, YOLOv3

黄文斌，陈仁文，袁婷婷. 改进YOLOv3-SPP的无人机目标检测模型压缩方案[J]. 计算机工程与应用, 2021, 57(21): 165-173.

HUANG Wenbin, CHEN Renwen, YUAN Tingting. Compression of UAV Object Detection Model Based on Improved YOLOv3-SPP[J]. Computer Engineering and Applications, 2021, 57(21): 165-173.

参考文献

[1] LECUN Y，BOTTOU L，BENGIO Y，et al.Gradient-based learning applied to document recognition[J].Proceedings of the IEEE，1998，86（11）：2278-2324.
[2] GIRSHICK R.Fast R-CNN[C]//IEEE International Conference on Computer Vision and Pattern Recognition，2015：1440-1448.
[3] REN S，HE K，GIRSHICK R B，et al.Faster R-CNN：towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2017，39（6）：1137-1149.
[4] DAI J，LI Y，HE K，et al.R-FCN：object detection via region-based fully convolutional networks[C]//Advances in Neural Information Processing Systems，2016：379-387.
[5] HE K，GKIOXARI G，DOLLAR P，et al.Mask R-CNN[C]//IEEE Conference on Computer Vision and Pattern Recognition，2018：6517-6525.
[6] REDMON J，DIVVALA S，GIRSHICK R，et al.You only look once：unified，real-time object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition，2016：6517-6525.
[7] LIU W，ANGUELOV D，ERHAN D，et al.SSD：single shot multibox detector[C]//Proceedings of European Conference on Commuter Vision，2016：21-37.
[8] REDMON J，FARHADI A.YOL09000：better，faster，stronger[C]//IEEE Conference on Computer Vision and Pattern Recognition，2017：6517-6525.
[9] REDMON J，FARHADI A.YOLOv3：an incremental improvement[J].arXiv：1804.02767，2018.
[10] BOCHKOVSKIY A，WANG C Y，LIAO H Y M.YOLOv4：optimal speed and accuracy of object detection[J].arXiv：2004.10934，2020.
[11] JOCHER G，STOKEN A，BOROVEC J，et al.Ultralytics/yolov5：v3.0（Version v3.0）[EB/OL]（2020-08-13）.http：//doi.org/10.5281/zenodo.3983579.
[12] 李云鹏，侯凌燕，王超.基于YOLOv2的复杂场景下车辆目标检测[J].电视技术，2018，42（5）：100-106.
LI Y P，HOU L Y WANG C.Vehicle object detection in complex scene based on yolov2[J].Video Engineering，2018，42（5）：100-106.
[13] HE Y H，ZHANG X Y，SUN J.Channel pruning for accelerating very deep neural networks[J].arXiv：1707. 06168，2017.
[14] HAN S，MAO H，DALLY W J.Deep compression：compressing deep neural networks with pruning，trained quantization and huffmancoding[EB/OL].[2015-10-15].https：//arxiv.org/abs/1510.00149.
[15] 李江昀，赵义凯，薛卓尔，等.深度神经网络模型压缩综述[J].工程科学学报，2019，41（10）：1229-1239.
LI J Y，ZHAO Y K，XUE Z E.A survey of model compression for deep neural networks[J].Chinese Journal of Engineering，2019，41（10）：1229-1239.
[16] LI H，KADAV A，DURDANOVIC I，et al.Pruning filters for efficient convnets[J].arXiv：1608.08710，2016.
[17] LIU Z，LI J，SHEN Z，et al.Learning efficient convolutional networks through network slimming[C]//2017 IEEE International Conference on Computer Vision（ICCV），2017.
[18] ZHANG P，ZHONG Y，LI X.SlimYOLOv3：narrower，faster and better for real-time UAV applications[C]//Proceedings of the IEEE International Conference on Computer Vision Workshops，2019.
[19] HE K，ZHANG X，REN S，et al.Spatial pyramid pooling in deep convolutional networks for visual recognition[J].IEEE Transactions on Pattern Analysis & Machine Intelligence，2014，37（9）：1904-1916.
[20] REZATOFIGHI H，TSOI N，GWAK J Y，et al.Generalized intersection over union：a metric and a loss for bounding box regression[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition（CVPR），2020.
[21] ZHONG Z，ZHENG L，KANG G，et al.Random erasing data augmentation[J].arXiv：1708.04896，2017.
[22] DEVRIES T，TAYLOR G W.Improved regularization of convolutional neural networks with cutout[J].arXiv：1708.04552，2017.
[23] EFRAIMIDIS P S，SPIRAKIS P G.Weighted random sampling with a reservoir[J].Information Processing Letters，2006，97（5）：181-185.
[24] LING C X，LI C H.Data mining for direct marketing：problems and solutions[C]//Proceedings of the Fourth International Conference on Knowledge Discovery and Data Ming，1998：73-79．
[25] GUO H X，LI Y J，JENNIFER S，et al.Learning from class-imbalanced data：review of methods and applications[J].Expert Systems with Applications，2017，73：220-239．
[26] TAKAHASHI R，MATSUBARA T，UEHARA K.Data augmentation using random image cropping and patching for deep CNNs[J].arXiv：1811.09030，2018.
[27] LIU Z，LI J，SHEN Z，et al.Learning efficient convolutional networks through network slimming[C]//Proceedings of the IEEE International Conference on Computer Vision，2017：2736-2744.