YOLO-U:基于结构重参数化和双重注意力机制的水下目标检测算法

YOLO-U: An underwater object detection algorithm based on structural reparameterization and dual attention mechanism

摘要:

水下目标检测算法的研究是实现水下机器人智能捕捞的前提。水下目标检测任务中存在的目标模糊、小目标众多以及相互遮挡等问题对实现精确的目标检测提出了挑战，本研究提出了一种基于YOLOv7-tiny的水下目标算法YOLO-U。该算法通过引入具有结构重参数化的RepViT骨干网络，融合ESE通道注意力机制，增强水下模糊目标的特征提取能力；同时，设计了浅层坐标信息特征融合的特征金字塔网络CAFPN，进一步增强检测模型对方向和位置信息的敏感度，融合不同尺度的特征信息以提高小目标的检测能力；最后，采用WIoUv2边界框损失函数有效降低了简单示例对损失值的贡献，使得模型能够聚焦于遮挡目标，进一步提高遮挡目标的检测精度。YOLO-U算法在URPC2021数据集上mAP₅₀取得了84.6%的检测效果，较YOLOv7-tiny、YOLOv5s和YOLOv8s分别提高了2.1%、5.2%和2.8%，检测结果显示，该算法可以有效提高水下目标检测精度，进一步改善水下模糊目标、小目标和遮挡目标的检测效果。

Abstract:

The research on underwater object detection algorithms is a prerequisite for achieving intelligent fishing with underwater robots. The problems of fuzzy object， numerous small objects and mutual occlusion in underwater object detection pose challenges to the realization of accurate object detection. This paper proposes a YOLO-U algorithm for underwater object detection based on YOLOv7-tiny. The algorithm introduces a RepViT backbone network with structural reparameterization and fuses an ESE channel attention mechanism to enhance the feature extraction capability for underwater fuzzy objects. Additionally， a feature pyramid network CAFPN with shallow coordinate information feature fusion is designed to further enhance the sensitivity of the detection model to directional and positional information， and integrate feature information of different scales to improve the detection ability of small objects. Furthermore， the WIoUv2 bounding box loss function is employed to effectively reduce the contribution of easy examples to the loss value. This allows the model to focus on occluded objects and further improve the detection accuracy for occluded objects. The YOLO-U algorithm achieves a mAP50 of 84.6% on the URPC2021 dataset， which is an improvement of 2.1%， 5.2%， and 2.8% compared to YOLOv7-tiny， YOLOv5s， and YOLOv8s， respectively. The detection results show that the algorithm can effectively improve the detection accuracy of underwater objects and further improve the detection performance of underwater fuzzy objects， small objects， and occluded objects.