改进YOLOv5的水面小目标检测算法  

Improved YOLOv5 Water Surface Small Target Detection Algorithm

作　　者：甘浪雄[1,2] 王梦颖 韩延胜 冯辉 GAN Langxiong;WANG Mengying;HAN Yansheng;FENG Hui(School of Navigation,Wuhan University of Technology,Wuhan 430063,China;Hubei Key Laboratory of Inland Shipping Technology,Wuhan 430063,China;Wuhan Culture&Tourism Group Co.Ltd.,Wuhan 430014,China;School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan 430063,China)

机构地区：[1]武汉理工大学航运学院,武汉430063 [2]内河航运技术湖北省重点实验室,武汉430063 [3]武汉文化旅游集团有限公司,武汉430014 [4]武汉理工大学船海与能源动力工程学院,武汉430063

出　　处：《武汉理工大学学报(交通科学与工程版)》2025年第2期448-454,共7页Journal of Wuhan University of Technology(Transportation Science & Engineering)

摘　　要：本文提出了一种改进YOLOv5的水面小目标检测算法.在网络结构上对浅层特征进行融合,新增一个检测头用于微小目标的检测.利用ConvMixer的结构特性,设计C3_CML模块用于取代原主干网络和颈部网络中特定位置的C3模块,通过增强图像特征信息空间通道位置关系的提取能力,从而提升对有效目标区域的关注,同时降低模型复杂度.设计了新的损失函数,综合使用IOU(intersection over union)和NWD(normalized wasserstein distance)作为新的边界框损失评价指标,降低对小目标位置偏差的敏感性,显著提高小目标的检测性能.结果表明:相比原始YOLOv5算法,改进后的算法有效减少了水面密集小目标和极小目标的漏检率,同时检测精度得到了显著提高.An improved YOLOv5 algorithm for detecting small targets on water surface was proposed.The shallow features were fused in the network structure,and a detection head was added to detect ti-ny targets.Using the structural characteristics of ConvMixer,C3_CML module was designed to re-place C3 module in specific position in the original backbone network and neck network.By enhancing the ability to extract the spatial channel position relationship of image feature information,the atten-tion to the effective target area was improved and the complexity of the model is reduced.A new loss function was designed,and IOU(intersection over union)and NWD(Normalized Wasserstein Dis-tance)were used as new loss evaluation indexes of the bounding box,which reduced the sensitivity to the position deviation of small targets and significantly improves the detection performance of small targets.The results show that,compared with the original YOLOv5 algorithm,the improved algo-rithm effectively reduces the missed detection rate of dense small targets and minimal targets on the water surface,and the detection accuracy is significantly improved.

关 键 词：水面小目标检测 YOLOv5 ConvMixer NWD 

分 类 号：U692.2[交通运输工程—港口、海岸及近海工程]