基于DR-YOLOv11的轻量化水下模糊目标检测算法

Detection Algorithm of Lightweight Underwater Blurry Object Based on DR-YOLOv11

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，提出了一种改进YOLOv11n的水下场景增强型检测算法DR-YOLOv11；该算法引入RFAConv替换标准卷积，通过动态生成感受野内的特征权重，提升局部特征提取能力；引入RFCAConv嵌入C3K2模块，融合感受野注意力与坐标注意力，增强特征提取的鲁棒性；设计了一种上下文引导的局部位置注意力机制CGLPA，通过构建上下文先验并融合通道特征信息，引导邻域特征自适应加权，提升模型对模糊环境下小目标的检测效果；采用BiFPN重构颈部结构，在降低模型复杂度的同时实现特征融合；通过DUO水下目标检测数据集上的实验表明，DR-YOLOv11的mAP@0.5与mAP@0.5:0.95分别达到86.0 %和66.5 %，较YOLOv11n提升3.5与3.3个百分点；同时模型参数量仅1.8 M，较基线缩减31 %。

Abstract:

To address the problem of degraded detection accuracy caused by target blurring and low edge contrast in underwater environments, this paper proposes DR-YOLOv11, an enhanced underwater scene detection algorithm based on YOLOv11n. The algorithm introduces RFAConv to replace standard convolutions, dynamically generating feature weights within the receptive field to enhance local feature extraction capability. RFCAConv is introduced into the C3K2 module, integrating receptive-field attention with coordinate attention to enhance the robustness of feature extraction. Furthermore, a Context-Guided Local Position Attention mechanism (CGLPA) is designed, which constructs contextual priors and fuses channel feature information to guide adaptive weighting of neighborhood features, thereby improving the detection performance for small targets in blurred environments. Bi-directional Feature Pyramid Network (BiFPN) is adopted to reconstruct the neck structure, achieving feature fusion while reducing model complexity. Experiments on the DUO underwater object detection dataset show that DR-YOLOv11 achieves 86.0 % mAP@0.5 and 66.5 % mAP@0.5:0.95 with only 1.8 M parameters, surpassing YOLOv11n by 3.5 and 3.3 percentage points, respectively, while reducing the parameter count by 31 %.