基于UW-YOLOv8的水下模糊小目标检测算法

UW-YOLO: An Algorithm for Underwater Small Object Detection in Blurry Environments

摘要:

针对水下目标模糊问题，提出了一种基于改进的YOLOv8的水下目标检测算法UW-YOLOv8；该算法引入了EfficientNetV1作为主干网络，EfficientNetV1通过均衡地缩放网络的深度、宽度和分辨率，提高了特征提取能力；在颈部增加了HAT注意力机制，HAT结合了通道注意力机制，增强了像素间的关系和特征交互，有效处理低分辨率水下图像；通过方法ASFF创新的检测头，并结合额外增加的小目标检测层，实现了不同层级特征图特征的融合，从而更细致提取小目标的特征，提高了模型对小尺度目标的检测精度。通过在URPC2020数据集上挑选的6000张模糊图片的实验结果表明，UW-YOLOv8相对于YOLOv8的mAP50和mAP50-95分别提升了2.8%和3%。

Abstract:

Addressing the issue of underwater target ambiguity, this paper proposes UW-YOLOv8, an underwater target detection algorithm based on an improved YOLOv8. This algorithm incorporates EfficientNetV1 as the backbone network. By scaling the depth, width, and resolution of the network in a balanced manner, EfficientNetV1 enhances feature extraction capabilities. Additionally, the Hybrid Attention (HAT) mechanism is introduced in the neck part, which integrates channel attention mechanisms to strengthen the relationships between pixels and feature interactions, effectively tackling low-resolution underwater images. The innovative Adaptive Spatial Feature Fusion (ASFF) method is employed in the detection head, combined with an extra small target detection layer, enabling the fusion of features from different hierarchical feature maps. This facilitates more detailed extraction of small target features and improves the detection accuracy for small-scale targets. Experimental results based on 6,000 blurred images selected from the URPC2020 dataset demonstrate that UW-YOLOv8 achieves improvements of 2.8% and 3% in mAP50 and mAP50-95, respectively, compared to YOLOv8.