改进YOLOv8M的水下目标检测算法  

Improved Underwater Target Detection Algorithm of YOLOv8M

作　　者：全文君 骆红涛 QUAN Wen-jun;LUO Hong-tao(School of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《计算机技术与发展》2025年第1期30-37,共8页Computer Technology and Development

基　　金：重庆理工大学科研启动基金(0103190032)。

摘　　要：受限于水中检测设备内存欠缺和计算能力不够强大以及水下环境错综复杂等多类因素,水下目标检测会出现检测目标错误和漏检目标等问题。为了解决上述问题,通过对YOLOv8M算法进行改进,提出了一种更高效、精准的算法,使水下目标检测检测性能得到了一定提升。首先,将一种注意力机制(A Simple-Atten Module,SimAM)嵌入网络结构中的颈部,解决了现有注意力模块只能沿着通道或空间维度细化特征,限制学习跨通道和空间变化的注意力权重灵活性的问题,加快了权重计算,提高了模型在水下模糊环境下的检测精度;其次,采用DCNv4(Deformable Convolution v4)可变形卷积替换C2f结构中的3×3卷积模块,以增强空间聚合的动态特性和表达能力并优化内存访问提高速度。此外,还采用加权双向特征金字塔结构(Bidirectional Feature Pyramid Network,BiFPN)将深度的特征信息层进行多尺度特征融合,以增强骨干网络的特征提取能力;最后,在预测部分使用Wise-Inner-MPDIoU损失函数将交叉熵损失函数进行替代,使模型对不同目标及尺寸的自适应能力得到一定程度的提升,增强算法的鲁棒性。从实验结果可以看出,模型经过改进后在URPC2020数据集上取得了86.4的mAP,相比于原始的YOLOv8M算法,mAP提高了2.8百分点,检测速度提升至102.9 FPS。经过改进后算法模型在进行水下目标检测任务时,无论是检测的精度或者速度都得到了有效的改进,说明该算法在用于水下目标检测的实际场景中有一定的实用性。Due to many factors such as the lack of memory and computational capabilities of underwater detection equipment,as well as the intricacies of the underwater environment,there will be problems such as detecting target error and missing target in underwater detection.In order to solve the above problems,we propose a more efficient and accurate algorithm by improving the YOLOv8M algorithm,which enhances the performance of underwater target detection to some extent.Firstly,SimAM(A Simple-Atten Module),an attention mechanism,is embedded in the neck of the network structure,which solves the problem that the existing attention module can only narrow features along the channel or spatial dimension,limiting the flexibility of learning attention weights that change across channels and spaces,speeds up the weight calculation,and improves the detection accuracy of the model in underwater fuzzy environment.Secondly,Deformable Convolution v4(DCNv4)is used to replace 3×3 convolution modules in the C2f structure to enhance the dynamic characteristics and expression ability of spatial aggregation and optimize memory access to improve speed.In addition,the weighted BiFPN(Bidirectional Feature Pyramid Network)is used to carry out multi-scale feature fusion of deep feature information layers to enhance the feature extraction capability of the backbone network.Finally,Wise-Inner-MPDIoU loss function is used to replace the cross-entropy loss function in the prediction part,so that the adaptive ability of the model to different targets and sizes is improved to a certain extent,and the robustness of the algorithm is enhanced.As can be seen from the experimental results,the improved model has obtained a mAP of 86.4 on the URPC2020 data set.Compared with the original YOLOv8M algorithm,the mAP has increased by 2.8 percentage points,and the detection speed has increased to 102.9 FPS.When the improved algorithm model is used for underwater target detection,both the detection accuracy and speed are effectively improved,indicating that the al

关 键 词：YOLOv8M 水下目标检测 注意力机制 金字塔结构 可变形卷积 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]