融合协同注意力机制与Transformer模型的鱼类异常行为多任务识别  

Fish Abnormal Behavior Multi-Task Recognition Based on Fusion of Collaborative Attention Mechanism and Transformer Mode

作　　者：张艺爔 胡泽元 左宇琪 贾松怡 刘吉航 陶红希 于红 ZHANG Yixi;HU Zeyuan;ZUO Yuqi;JIA Songyi;LIU Jihang;TAO Hongxi;YU Hong(College of Information Engineering,Dalian Ocean University;Dalian Key Laboratory of Smart Fishery;Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University),Ministry of Education;Key Laboratory of Marine Information Technology of Liaoning Province,Dalian 116023,China)

机构地区：[1]大连海洋大学信息工程学院 [2]大连市智慧渔业重点实验室 [3]设施渔业教育部重点实验室(大连海洋大学) [4]辽宁省海洋信息技术重点实验室,辽宁大连116023

出　　处：《广东海洋大学学报》2025年第1期124-133,共10页Journal of Guangdong Ocean University

基　　金：辽宁省科技计划联合基金(2023-BSBA-001);辽宁省教育厅重点科研项目(LJKZ0729);海洋渔业领域知识图谱构建项目(2023JH26/10200015);辽宁省教育厅基本科研项目(JYTQN2023132);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：【目的】解决暗光、浑浊和高密度养殖环境下,对于单任务鱼类异常行为识别精确度不高以及相似性鱼类异常行为难以准确识别的问题。【方法】提出一种多任务学习情况下鱼类异常行为及其姿态估计研究框架,命名为PD-DETR。通过Transformer架构实现端到端推理,平衡不同任务的损失权重,优化梯度冲突。通过自注意力编码器和协同注意力(SCSA)特征融合网络(SCSA-FPN),计算单鱼行为与鱼群行为的权重,平衡鱼群行为对个体行为的影响,降低相似性行为特征丢失。设计消融实验和模型对比实验,以证算法的有效性。【结果】PD-DETR在红鳍东方鲀(Takifugu rubripes)异常行为数据集上的识别精确率和平均精度分别达到95.1%和93.6%,较YOLOv11-det提升0.9%和0.3%;游动姿态估计精确率和平均精度分别达到91.2%和90.8%,较RT-DETR相比分别提升3.9%和4.4%;在多任务学习情况下异常识别任务和游动姿态估计任务的平均精度较单任务学习提升1.2%和1.7%。【结论】多任务学习网络PD-DETR实现了暗光、浑浊水质环境中的鱼类异常行为识别与游动姿态分析,有助于提高养殖效率,保障鱼类健康。【Objective】To solve the problem of low accuracy in identifying abnormal behavior of single task fish and difficulty in accurately identifying abnormal behavior of similar fish in dimlight,turbid,and high-density aquaculture environments.【Method】A research framework named PD-DETR was proposed for estimating abnormal behavior and posture of fish under multi-task learning.Transformer architecture was applied for implementing end-to-end inference balancing the loss weights of different tasks,and optimizing gradient conflicts.By using self-attention encoder and collaborative attention feature fusion network(SCSA-FPN),the weights of single fish behavior and fish shoal behavior were calculated to balance the impact of fish shoal behavior on individual behavior and reduce the problem of loss of similar behavior features.Ablation experiments and model comparison experiments were designed to demonstrate the effectiveness of the algorithm.【Result】The recognition accuracy and average accuracy of PD-DETR on Takifugu rubripes abnormal behavior dataset reached 95.1%and 93.6%,respectively,which were 0.9%and 0.3%higher than YOLOv11 det.The accuracy and average accuracy of swimming pose estimation reached 91.2%and 90.8%,respectively,which were 3.9%and 4.4%higher than RT-DETR.In the case of multi-task learning,the average accuracy of anomaly recognition and swimming pose estimation was 1.2%and 1.7%respectively higher than that of single task learning.【Conclusion】The multi-task learning network PD-DETR in this study can perform abnormal behavior recognition and swimming posture analysis of fish in dimlight and turbid water environments,which helps to improve aquaculture efficiency and ensure fish welfare.

关 键 词：鱼类异常行为 姿态估计 多任务学习 多任务梯度协调 Transformer模型 

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