水下复杂环境高效鲁棒目标检测方法

Research on efficient and robust object detection of complex underwater environments

针对水下环境复杂多变且视觉退化严重,现有的水下目标检测算法难以保证高精度实时检测水下目标且在复杂环境中的检测鲁棒性不足等问题,本文的算法架构基于YOLOv5改进,根据通用模型提出一种端到端的多尺度水下目标检测网络算法(UW-Net),完成复杂水下环境中高精度实时鲁棒检测水下目标任务。在特征提取部分,该网络通过稳定的底层特征提取模块和CSP-Net构建高精度轻量型特征提取结构,旨在保证网络实时性的同时提取更高维度的特征信息;在特征融合和检测部分,使用自适应特征融合机制和注意力增强方法在基本不影响检测速度的同时,提升算法的多尺度目标能力和检测鲁棒性,并通过K-means聚类方法自监督的实现最优锚框的标定从而实现目标区域的准确预知。实验结果表明:该方法在NVIDIA GeForce GTX1080Ti的GPU平台上对水下目标数据集的平均检测精度和检测速度分别为95.06%和139FPS,比YOLOv5s网络提升了2.87%和14FPS,实现了在实际复杂水下环境中高精度实时鲁棒地检测水下目标。

In view of the changing underwater environments and significant visual degradation, the existing underwater object detection algorithms are difficult to guarantee high-precision real-time detection of underwater objects and perform robustness detection in complex environments. This paper presents an end-to-end multi-scale underwater object detection network model (UW-Net) for detecting underwater objects in complex underwater environments with accuracy and efficiency. The network constructs a reliable and lightweight feature extraction structure through a stable underlying feature extraction module and CSP-Net, which aims to extract richer feature information during the real-time extraction. As part of the feature fusion and detection process, the adaptive feature fusion mechanism and the combined attention enhancement method are used to enhance the multi-scale object detection capability and the robustness of the algorithm without affecting its detection speed. Moreover, the self-supervised calibration of the optimal anchor frame is achieved through the K-means clustering method to achieve accurate prediction of the target area. The experimental results show that the detection accuracy and speed of the method on the underwater object dataset on GPU platform of NVIDIA GeForce GTX1080Ti are 95.06% and 139 FPS respectively, which are 2.87% and 14FPS higher than those of the YOLOv5s network and achieve the real-time and robust detection of underwater objects in complex environments with high-accuracy.