用于复杂水下环境下的轻量级目标检测算法  

Lightweight object detection algorithm for complex underwater environment

作　　者：王飞飞 秦玉芳[1,2] 冯国富 陈明[1,2] WANG Feifei;QIN Yufang;FENG Guofu;CHEN Ming(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China;Key Laboratory of Fisheries Information,Ministry of Agriculture and Rural Affairs,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]农业农村部渔业信息重点实验室,上海201306

出　　处：《传感器与微系统》2025年第4期137-142,147,共7页Transducer and Microsystem Technologies

基　　金：山东省重点研发计划(乡村振兴科技创新提振行动计划)项目(2023TZXD051);广东省重点领域研发计划项目(2021B0202070001)。

摘　　要：在利用水下机器人进行生物捕捞与识别时,色偏、雾效应与光干扰等复杂水下环境的干扰会导致模型存在漏检和误检现象,同时,由于带宽受限,水下设备计算资源也受限制。为解决上述问题,提出了一种高检测精度的轻量级改进的YOLOv7-tiny模型。使用SiLU激活函数和EIOU损失函数提高模型收敛速度;使用双重注意力叠加的GAM注意力机制和三重注意力叠加的DyHead检测头更好地抵御环境噪声的干扰;最后,利用通道剪枝策略降低模型参数量。实验结果表明:在RUOD数据集上,改进后的模型有效的缓解了色偏、雾效应与光干扰等复杂水下环境下的漏检和误检现象,浮点数计算量降低8.4%,提高了水下设备在计算资源受到限制时的可用性。所提方法可适用于水下检测系统和水下机器人。When using underwater robots for biological fishing and recognition,the interference of complex underwater environment,such as color bias,fog effect and light interference,will lead to missing and false detection phenomena in model.Meanwhile,due to limited bandwidth,the computing resources of underwater equipment are also limited.To solve the above problems,a lightweight improved YOLOv7-tiny model with high detection precision is proposed.The SiLU activation function and EIOU loss function are used to improve the model convergence speed.The GAM attention mechanism with double attention overlayed and DyHead detection head with triple attention superposition can better resist the interference of environmental noise.Finally,channel pruning strategy is used to reduce the number of model parameters.The experimental results show that on the RUOD data set,the improved model effectively alleviates the phenomenon of missing and false detection in complex underwater environment such as color bias,fog effect and light interference,and the calculation amount of floating point number is reduced by 8.4%,which improves the availability of underwater equipment when computing resources are limited.The proposed method can be applied to underwater detection system and underwater robot.

关 键 词：水下生物目标检测 YOLOv7-tiny 注意力机制 通道剪枝 复杂水下环境 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]