基于YOLOv7-tiny的轻量化海珍品检测算法  

Lightweight sea treasure detection algorithm based on YOLOv7-tiny

作　　者：陈俊逸 曹立杰 吴军[1] 罗佳璐 何植仟 CHEN Junyi;CAO Lijie;WU Jun;LUO Jialu;HE Zhiqian(College of Information Engineering,Dalian Ocean University,Dalian Liaoning 116023,China;Liaoning Provincial Key Laboratory of Marine Information Technology(Dalian Ocean University),Dalian Liaoning 116023,China)

机构地区：[1]大连海洋大学信息工程学院,辽宁大连116023 [2]辽宁省海洋信息技术重点实验室(大连海洋大学),辽宁大连116023

出　　处：《计算机应用》2024年第S01期319-323,共5页journal of Computer Applications

基　　金：辽宁省教育厅科研经费资助项目(LJKZ0731)。

摘　　要：针对当前海珍品捕捞机器人使用的水下目标检测算法参数量大,不适合部署在移动设备上等问题,提出一种基于YOLOv7-tiny(You Only Look Once version 7-tiny)的轻量化海珍品检测算法ES YOLOv7-tiny(EfficientNet-S YOLOv7-tiny)。在YOLOv7-tiny基础上,首先,将骨干网络替换为改进的EfficientNet(EfficientNet-S),并将颈部网络中卷积核大小为3×3卷积替换为轻量化卷积,达到降低参数量的目的;其次,使用k-means++算法聚类锚框尺寸,提高推理速度;最后,使用知识蒸馏算法进一步提高精度。在RUIE(Real-world Underwater Image Enhancement)数据集上,所提算法平均精度均值(mAP)达到73.7%,检测速度达到123 frame/s,参数量为4.45×10^(6),与原YOLOv7-tiny算法相比,在mAP上提升了1.2个百分点,检测速度提升25 frame/s,参数量降低了1.56×10^(6)。实验结果表明,所提算法在提升精度的同时降低了参数量,并且加快了检测速度,证明了该算法的有效性。Because the current underwater target detection algorithms used in sea treasure fishing robots have too large number of parameters to deploy on mobile devices,a lightweight sea treasure detection algorithm ES YOLOv7-tiny(EfficientNet-S You Only Look Once version 7-tiny)based on YOLOv7-tiny(You Only Look Once version 7-tiny)was proposed.On the basis of YOLOv7-tiny,first,the backbone network was replaced with an improved EfficientNet(EfficientNet-S),and the convolution kernel size of 3×3 in the neck network was replaced with lightweight convolution to reduce the number of parameters.Secondly,k-means++algorithm was used to cluster the anchor box sizes to improve inference speed.Finally,knowledge distillation algorithm was used to further improve accuracy.On the RUIE(Real-world Underwater Image Enhancement)dataset,the mean Average Precision(mAP)of proposed algorithm reaches 73.7%,with a detection speed of 123 frame/s and a parameter volume of 4.45×10^(6).In comparison with the original YOLOv7-tiny algorithm,ES YOLOv7-tiny algorithm improved mAP by 1.2 percentage points,increased detection speed by 25 frame/s,and decreased parameter volume by 1.56×10^(6).The experimental results show that proposed algorithm reduces the number of parameters while improving accuracy,and accelerates detection speed,proving the effectiveness of the algorithm.

关 键 词：海珍品 目标检测 YOLOv7-tiny 轻量化 k-means++ 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]