基于DMS-YOLO的水下模糊目标检测算法

DMS-YOLO: An Underwater Blurred Object Detection Algorithm

摘要: 针对水下目标模糊、边缘对比度低导致检测精度下降的问题，本文提出了一种改进的水下目标检测模型DMS-YOLO。该方法首先引入动态上采样模块(Dynamic Sampling, DySample)，通过自适应调整采样位置实现特征重建，以改善细节信息的恢复效果；其次构建动态检测头(Dynamic Head, DyHead)，从尺度、空间及任务三个维度提升特征表达的自适应性；进一步在C2PSA结构中嵌入多尺度膨胀注意力机制(Multi-Scale Dilated Attention, MSDA)，以增强多尺度信息融合能力；最后采用Slide Loss损失函数，通过动态加权策略提升模型对分类困难样本的学习能力。在RUOD数据集上的实验结果表明，DMS-YOLO的mAP@0.5达到87.5%，mAP@0.5:0.95为64.0%，该方法对模糊目标细节表达及多尺度特征融合方面表现出一定改进，并在检测精度与计算复杂度之间呈现出较为合理的平衡。

Abstract: To address the decline in detection accuracy caused by blurred underwater targets and low edge contrast, this paper proposes an improved underwater object detection model, DMS-YOLO. The method first introduces a dynamic upsampling module (Dynamic Sampling, DySample), which reconstructs features by adaptively adjusting sampling positions to improve the recovery of fine-grained details. Next, a dynamic detection head (Dynamic Head, DyHead) is constructed to enhance the adaptability of feature representation from three dimensions: scale, spatial, and task. Furthermore, a multi-scale dilated attention mechanism (Multi-Scale Dilated Attention, MSDA) is embedded into the C2PSA structure to strengthen multi-scale feature fusion. Finally, the Slide Loss function is adopted, which improves the model’s learning capability for difficult classification samples through a dynamic weighting strategy. Experimental results on the RUOD dataset show that DMS-YOLO achieves an mAP@0.5 of 87.5% and an mAP@0.5:0.95 of 64.0%. The results indicate that the proposed method shows improvements in the representation of blurred target details and multi-scale feature fusion, while maintaining a relatively balanced trade-off between detection accuracy and computational complexity.