基于SimAM注意力机制的DCN-YOLOv5水下目标检测  

DCN-YOLOv5 Underwater Target Detection Based on SimAM Attention Mechanism

作　　者：刘向举[1] 刘洋 蒋社想[1] LIU Xiangju;LIU Yang;JIANG Shexiang(School of Computer Science and Engineering,Anhui University of Science and Technology,Anhui Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《重庆工商大学学报(自然科学版)》2025年第2期63-70,共8页Journal of Chongqing Technology and Business University:Natural Science Edition

基　　金：安徽省重点实验室项目(ZKSYS202204).

摘　　要：目的针对水下环境复杂,水下目标因光线折射导致的目标边界模糊或外观、形状可能会发生非刚性形变,使水下目标检测困难的问题,提出了一种基于SimAM注意力机制的DCN-YOLOv5水下目标检测方法。方法首先,采用YOLOv5所使用的双向金字塔网络(BiFPN,Bi-directional Feature Pyramid Network)在多个尺度上提取和融合特征信息,从而提高目标辨别的准确度;其次,针对水下目标的外观、形状变化问题,将C3模块中的CBS模块结合可变形卷积(DCN,Deformable Convolution Network),提出DBS模块并组成D3模块替换部分C3模块,以适应水下目标的外观、形状变化;同时,融入加权注意力机制(SimAM),自适应地调节模型的关注度,进一步在复杂场景下增强特征表达能力;最后,考虑目标边界模糊,为改善目标定位精度,采用WIoU(Wise-IoU)损失函数来替换交叉熵损失,能够更好地适应不同目标类型和尺寸的特点,提高算法鲁棒性。结果实验结果表明:DCN-YOLOv5可以达到87.57%的平均精度(mAP),检测效果优于YOLOv5网络和其他经典网络,平均每张图像的识别时间仅为24.5 ms。结论通过实验结果可以证明模型在检测精度明显提升的同时兼顾检测的实时性,对水下目标检测用于实际用途有着一定的参考价值。Objective Given the complex underwater environment,the target boundary may be blurred or the appearance and shape of the underwater target may be non-rigidly deformed due to light refraction,which makes underwater target detection difficult.A DCN-YOLOv5 underwater target detection method based on the SimAM attention mechanism was proposed.Methods Firstly,the bi-directional feature pyramid network(BiFPN)used by YOLOv5 was used to extract and fuse feature information on multiple scales to improve the accuracy of target recognition.Secondly,to address the variations in appearance and shape of underwater objects,the CBS module in the C3 module was combined with the deformable convolution network(DCN),and the DBS module was proposed.The DBS module was used to form the D3 module and replace part of the C3 module to adapt to the changing appearance and shape of the underwater targets.At the same time,the weighted attention mechanism was integrated to adaptively adjust the attention of the model and further improve the feature expression ability in complex scenes.Finally,considering the fuzzy boundary of the target and to improve the target positioning accuracy,the WIoU(Wise-IoU)loss function was used to replace the cross-entropy loss,which can better adapt to the characteristics of different target types and sizes and improve the robustness of the algorithm.Results Experimental results showed that DCN-YOLOv5 achieved an average precision(mAP)of 87.57%,outperforming YOLOv5 and other classical networks,with an average identification time of only 24.5 ms per image.Conclusion The experimental results demonstrate that the model significantly improves detection accuracy while ensuring real-time detection,providing valuable insights for the practical use of underwater target detection.

关 键 词：水下目标检测 SimAM注意力机制 可变形卷积 WIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]