基于深度学习的复杂环境下船舶视觉检测方法  

Deep Learning-Based Ship Detection Approach under Complex Environments

作　　者：崔雨箫 陈超 潘宝峰 叶翔 CUI Yuxiao;CHEN Chao;PAN Baofeng;YE Xiang(School of Naval Architecture and Maritime,Zhejiang Ocean University,Zhoushan 316000,Zhejiang,China)

机构地区：[1]浙江海洋大学船舶与海运学院,浙江舟山316000

出　　处：《船舶工程》2024年第2期106-114,共9页Ship Engineering

基　　金：浙江省“尖兵”“领雁”研发攻关计划(2023C03181);企业行业难题攻关项目(1118106412204,1118106412301)。

摘　　要：为减少由雷达盲点、船舶自动识别系统(AIS)信息不可靠以及人工疏忽等因素引起的海上安全事故,结合深度学习的计算机视觉技术,提出一种基于改进YOLOv7算法的船舶视觉检测方法(YOLOv7-Rep VGG-SimAM)。该算法将卷积核的每个通道进行分离,利用Rep VGG网络替代标准的卷积层以减少计算量;将Sim AM无参数注意力机制融合能量函数计算注意力权重,增强了对多船重叠和背景遮挡等条件下船舶特征的提取能力;采用暗通道原理进行去雾预处理,提高了检测的准确性和稳定性。试验结果表明:与传统的YOLOv5和YOLOv7等方法相比,YOLOv7-RepVGG-SimAM算法在检测精度、准确率和鲁棒性等方面均表现良好,尤其是复杂环境下,其船舶检测率依然维持在较高水平。To reduce maritime safety accidents caused by radar blind spots,unreliable automatic identification system(AIS)information,and human negligence,a ship visual detection method based on improved YOLOv7algorithm(YOLOV7-Repv GG-SIMAM)is proposed in combination with deep learning computer vision technology.This algorithm separates each channel of the convolutional kernel and replaces the standard convolutional layer with a Rep VGG network to reduce computational complexity and parameter count.It integrates the Sim AM mechanism to calculate attention weights using channel-wise similarities,thereby enhancing the ability to extract ship features under complex conditions such as multiple ship overlaps and background occlusion.Additionally,the algorithm uses the dark channel prior for haze removal preprocessing to improve detection accuracy and stability.Experimental results demonstrate that compared with traditional methods such as YOLOv5 and YOLOv7,the algorithm proposed in this paper exhibits superior performance in terms of detection accuracy,precision,and robustness,especially in complex environments when ships are used as detection targets,and the detection rate remains at a high level.

关 键 词：船舶视觉检测 深度学习 YOLOv7 船舶特征提取 RepVGG-SimAM 

分 类 号：U665.261[交通运输工程—船舶及航道工程]