基于通道非降维与空间协调注意力的改进YOLOv8养殖鱼群检测  

Farmed fish detection by improved YOLOv8 based on channel non-degradation with spatially coordinated attention

作　　者：涂万 于红 张鹏 韦思学[1,2] 张鑫 杨宗轶 吴俊峰[1,2] 林远山 胡泽元 TU Wan;YU Hong;ZHANG Peng;WEI Sixue;ZHANG Xin;YANG Zongyi;WU Junfeng;LIN Yuanshan;HU Zeyuan(Key Laboratory of Marine Information Technology of Liaoning Province,College of Information Engineering,Dalian Ocean University,Dalian 116023,China;Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University),Ministry of Education,Dalian 116023,China)

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

出　　处：《大连海洋大学学报》2023年第4期717-725,共9页Journal of Dalian Ocean University

基　　金：辽宁省教育厅重点科研项目(LJKZ0729);国家自然科学基金(31972846);设施渔业教育部重点实验室(大连海洋大学)开放课题(202313)。

摘　　要：为解决真实养殖环境中因鱼群模糊、遮挡造成的鱼类目标检测困难等问题,采用基于通道非降维与空间协调注意力ECAM(efficient coordination attention module)的改进YOLOv8养殖鱼群检测方法YOLOv8-Head-ECAM以提高检测精度。首先在FPN(feature pyramid network)中增加大尺寸检测头,更好地捕捉水下鱼类个体的细节信息,以加强对鱼群特征的提取能力,然后使用ECAM注意力机制减少模糊背景的干扰,聚焦鱼类个体的关键特征,以加强对模糊鱼群的识别能力,并设计了消融试验和模型对比试验以验证算法的有效性。结果表明:相比于YOLOv8,YOLOv8-Head-ECAM模型的准确率、召回率和平均精度均值分别提高了2.3%、1.7%和1.6%;与目前检测准确率较高的养殖鱼群检测模型KAYOLO、DCM-ATM-YOLOv5、SK-YOLOv5和ESB-YOLO相比,平均精度均值分别提高了0.7%、1.0%、2.4%和2.0%。研究表明,本文中提出的YOLOv8-Head-ECAM模型能够较好地适应水下鱼群模糊、遮挡的情况,提高了鱼群检测的有效性。In order to solve the problem of difficult fish target detection caused by fuzzy and occluded fish in actual aquaculture environments,YOLOv8-Head-ECAM,an improved YOLOv8 farmed fish detection method was established based on channel non-degradation with spatially coordinated attention ECAM(efficient coordination attention module).Firstly,the FPN(feature pyramid network)increases the large size detection head to better capture the detail information on underwater fish individuals to enhance the feature extraction ability of fish schools,and then the ECAM attention mechanism is used to reduce the interference of the fuzzy background and focuses on the key features of the fish individuals to enhance the recognition ability of fuzzy fish schools.The ablation test and comparison test were designed and carried out to verify the effectiveness of the algorithm.The results showed that the YOLOv8-Head-ECAM was improved by 2.3%in accuracy,by 1.7%in recall,and by 1.6%in mean average precision compared with the YOLOv8 model.There was 0.7%higher detection accuracy than the current farmed fish population detection models KAYOLO,1.0%higher detection accuracy than DCM-ATM-YOLOv5,2.4%higher detection accuracy than SK-YOLOv5,and 2.0%higher detection accuracy than ESB-YOLO.It is shown that the YOLOv8-Head-ECAM model proposed in this paper better adapts to the underwater fish blurring and occlusion and to satisfies the effectiveness of farmed fish detection.

关 键 词：注意力机制 YOLOv8 特征提取 鱼群检测 计算机视觉 

分 类 号：S932.2[农业科学—渔业资源] TP391[农业科学—水产科学]