摘要
为了在SAR图像中准确地检测出大范围海域内的舰船等目标,提出了一种基于YOLOv5-Transformer的目标检测算法:YOLOv5-TVC。首先,使用Vo VNet-57替代原有的CSP-Darknet53以增强对小目标特征的敏感性。其次,将CBAM加在neck的多尺度采样层中,实现深层网络中对重要空间和通道的关注。最后,在Bottleneck-Transformer模块内,用自我注意力机制叠加结构来替代提取特征的卷积层,从而优化对目标特征提取的效率。消融实验和对比实验的结果表明,YOLOv5-TVC检测SAR图像目标的精度优于其他YOLO系列算法。
An improved target detection algorithm based on Yolov5-transformer,YOLOv5-TVC,is proposed to precisely identify ships in a large area from SAR(Synthetic Aperture Radar)images.Firstly,the original CSP-DARKNET53 in YOLOv5 is replaced by VoVNet-57 to increase the sensitivity of small target features.Secondly,a convolutional block attention module(CBAM)is added to the multi-scale sampling layer of neck to realize the attention of important space and channels in deep networks.Finally,in the Bottleneck-Transformer module,the self-attention stack structure is used to replace the feature extraction convolution layer for optimizing efficiency of target feature extraction.The results of ablation experiments and contrast experiments show that the proposed YOLOv5-TVC has a better target detection performance than other YOLO series in precision.
作者
张翰康
颜明重
朱大奇
陈斌
李杰
ZHANG Hankang;YAN Mingzhong;ZHU Daqi;CHEN Bin;LI Jie(Logistics Engineering School,Shanghai Maritime University,Shanghai 201306,China;School of Mechanical Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Southwest Electronics and Telecommunication Technology Research Institute,Shanghai Branch,Shanghai 200434,China;Troops No.91746,China)
出处
《控制工程》
CSCD
北大核心
2023年第11期1979-1989,共11页
Control Engineering of China
基金
国家自然科学基金重点项目(62033009,U1706224)
上海市科技创新行动计划项目(206510712300,18DZ2253100)。