基于注意力机制及Ghost-YOLOv5的水下垃圾目标检测

DETECTION OF UNDERWATER TRASH BASED ON Ghost-YOLOv5 ANDATTENTION MECHANISM

水下垃圾的目标检测技术对水下机器人实现垃圾自动清除有着重要意义。然而,复杂的水下环境和水底光线不足,易导致检测精度受限、计算量大等问题。针对这些问题,提出了一种基于YOLOv5的水下垃圾目标检测的改进算法。在该方法中,在预处理部分引入Gamma变换提高水下图像的灰度和对比度,便于模型检测。同时,在YOLOv5检测部分嵌入CBAM注意力机制,以突出目标特征并抑制次要信息,从而提高算法精度。此外,将颈部层中的普通卷积模块替换为Ghost卷积模块,减少计算量,加快检测速度。采用真实环境下的水下垃圾数据集进行模型验证,与当前热门的目标检测算法进行对比,该方法在分辨率为640×640的图像上的最高检测精度为93.7%,且计算时间仅为6.7 ms,满足实时性的要求。该研究成果对水下垃圾的目标检测具有良好的借鉴意义。

Underwater trash detection technology is of great significance for automatic trash removal tasks by underwater robots.However,it faces some challenges,such as an unsatisfactory detection rate due to poor underwater light conditions and high computation load.To solve these problems,this paper proposed an improved YOLOv5 model for underwater trash detection.First,in the preprocessing stage,Gamma transform was introduced to improve the grey level and contrast of underwater images for model detection.Meanwhile,the CBAM attention mechanism was embedded in the detection part of the YOLOv5 model to select the information important to underwater trash detection tasks and suppress uncritical information,thus improving the accuracy of the algorithm.Besides,in the neck layer,the traditional convolution module was replaced by the Ghost convolution module to reduce the calculation amount and improve the detection speed.The proposed model was evaluated on an underwater trash dataset in the real environment.Compared with mainstream object detection algorithms,the proposed model achieved the highest detection accuracy of 93.7%in images with a resolution of 640×640,and the calculation time was only 6.7 ms,meeting the requirements of real-time performance.The study results provide a good reference for underwater trash detection.