YOLO11-Swin：一种面向复杂水下环境的目标检测模型

YOLO11-Swin: A Target Detection Model for Complex Underwater Environments

摘要: 水下目标检测在海洋资源开发与生态环境监测中至关重要，但水下图像的低对比度、色彩失真及复杂背景干扰为精准检测带来巨大挑战。为克服传统方法在特征提取与小目标识别上的局限，本文提出一种深度融合Swin Transformer与YOLO11架构的新型检测模型(A Novel Detection Model with Deep Integration of Swin Transformer and YOLO11 Architectures, YOLO11-Swin)。该模型以Swin Transformer作为主干特征提取网络，利用其分层设计与滑动窗口自注意力机制，有效捕获图像的全局上下文依赖关系，增强对模糊、遮挡目标的表征能力。在特征融合阶段，本文设计了一种跨层特征聚合机制(Cross-layer Feature Aggregation, CFA)，通过全局池化与自适应权重计算，引导不同尺度特征图进行高效信息交互，以解决特征金字塔中的语义间隙与尺度不匹配问题。此外，在各级特征图输出端嵌入卷积注意力模块(Convolutional Block Attention Module, CBAM)，通过串行的通道与空间注意力子模块，自适应地优化特征响应，突出目标区域并抑制背景噪声。针对水下数据集正负样本不均衡的问题，模型采用Focal Loss作为分类损失函数，以聚焦困难样本的训练，提升模型收敛速度与稳健性。在URPC数据集上的实验结果表明，YOLO11-Swin的mAP@50达到75.54%，相比基线YOLO11模型显著提升9.42%。特别地，对小目标(如扇贝)的检测平均精度(AP)提升10.16%，召回率(Recall)提高4.55%，充分验证了所提模型在复杂水下环境下的有效性与先进性。

Abstract: Underwater object detection plays a crucial role in marine resource development and ecological environment monitoring. However, the low contrast, color distortion, and complex background interference of underwater images pose significant challenges to accurate detection. To overcome the limitations of traditional methods in feature extraction and small object recognition, this paper proposes a novel detection model with deep integration of Swin Transformer and YOLO11 architectures (referred to as YOLO11-Swin). This model adopts Swin Transformer as the backbone feature extraction network. Leveraging its hierarchical design and sliding window self-attention mechanism, it effectively captures the global contextual dependencies of images and enhances the representation capability for blurred and occluded objects. In the feature fusion stage, a Cross-layer Feature Aggregation (CFA) mechanism is designed. Through global pooling and adaptive weight calculation, it guides efficient information interaction among feature maps of different scales, thereby addressing the issues of semantic gaps and scale mismatches in the feature pyramid. Additionally, Convolutional Block Attention Module (CBAM) is embedded at the output end of feature maps at all levels. Via serial channel and spatial attention sub-modules, it adaptively optimizes feature responses, highlights object regions, and suppresses background noise. To tackle the problem of imbalanced positive and negative samples in underwater datasets, the model employs Focal Loss as the classification loss function. This focuses on the training of hard samples, improving the model’s convergence speed and robustness. Experimental results on the URPC dataset demonstrate that the mAP@50 of YOLO11-Swin reaches 75.54%, which is a significant increase of 9.42% compared to the baseline YOLO11 model. Specifically, the Average Precision (AP) for small objects (e.g., scallops) is improved by 10.16%, and the Recall is increased by 4.55%. These results fully verify the effectiveness and advancement of the proposed model in complex underwater environments.