文献类型：期刊文献

中文题名：基于改进YOLO v7的鲑鱼检测模型轻量化研究

英文题名：Lightweight Salmon Detection Model Based on Improved YOLO v7

作者：郑荣才[1];谭鼎文[1,2];徐青[2];陈大勇[1];元轲新[1]

机构：[1]南方海洋科学与工程广东省实验室(湛江),湛江524013;[2]广东海洋大学机械工程学院,湛江524088

年份：2024

卷号：55

期号：11

起止页码：132

中文期刊名：农业机械学报

外文期刊名：Transactions of the Chinese Society for Agricultural Machinery

收录：北大核心2023、CSTPCD、、EI(收录号：20244817420479)、Scopus、CSCD2023_2024、北大核心、CSCD

基金：国家重点研发计划项目(2022YFD2401201);广东省海洋经济发展(海洋六大产业)专项资金项目(GDNRC[2023]33);南方海洋科学与工程广东省实验室(湛江)项目(011Z23002);湛江湾实验室人才团队引进科研项目(ZJW-2023-05)。

语种：中文

中文关键词：深远海养殖;鲑鱼检测;YOLO v7;Stem模块;多尺度重参数化;卷积块注意力模块

外文关键词：deep-sea aquaculture;salmon detection;YOLO v7;Stem module;multi-directional reparameterization;convolutional block attention module

中文摘要：为实现水下复杂环境下鲑鱼的快速准确识别,提出一种基于YOLO v7轻量化的鲑鱼检测模型YOLO v7-CSMRep。首先,采用Stem模块合并Backbone层的前4个卷积操作,有效降低了模型计算量。其次,使用多尺度重参数化(Multi-directional reparameterization,MRep)模块替代YOLO v7的ELAN和ELAN-H模块,增强了单向特征提取能力,同时大幅减少参数量和计算量。最后,在Backbone层末端集成卷积块注意力模块(Convolutional block attention module,CBAM),提升网络空间和通道特征提取能力。试验结果表明,改进后模型内存占用量、参数量和计算量分别降低4.28%、5.29%、31.30%,F1值、mAP05分别提高0.5、0.7个百分点,分别达到93.1%、97.1%,帧率提高15.41%,达到140.8 f/s。对比YOLO v5s、YOLO v6s、YOLO v7、YOLO v7-tiny、YOLO v8s模型,mAP05分别提高1.0、2.0、0.7、0.8、1.2个百分点。因此,本文提出的方法能够快速而准确地识别鲑鱼,可为深远海养殖生物量监测提供技术支撑。

外文摘要：In order to achieve rapid and accurate identification of salmon in complex underwater environments,a lightweight salmon detection model,YOLO v7-CSMRep,was proposed based on YOLO v7.Firstly,by adopting the Stem module,the first four convolutional operations in the backbone layer were merged into an efficient convolutional operation,reducing the computational load of the model.Secondly,the ELAN and ELAN-H modules of the YOLO v7 network were replaced with the multi-directional reparameterization(MRep)module,which enhanced the one-way feature extraction capability while greatly reducing parameters and calculations.Finally,at the end of the backbone layer,the convolutional block attention module(CBAM)was integrated to enhance the network’s spatial and channel feature extraction capabilities.The experimental results showed that the improved model’s volume,parameter count,and computational load were reduced by 4.28%,5.29%and 31.30%,respectively.The F1 score and mAP0.5 were increased by 0.5 and 0.7 percentage points,and reached 93.1%and 97.1%,respectively.Additionally,the frame rate was increased by 15.41%,and reached 140.8f/s.Compared with that of YOLO v5s,YOLO v6s,YOLO v7,YOLO v7-tiny,and YOLO v8s models,the mAP0.5 was improved by 1.0,2.0,0.7,0.8,and 1.2 percentage points,respectively.Therefore,the method proposed can rapidly and accurately identify salmon and provide technical support for biomass monitoring in deep-sea aquaculture.