旋转矩形框与CBAM改进RetinaNet的SAR图像近岸舰船检测

SAR Images Nearshore Ship Detection Based on RetinaNet Algorithm with Rotated Rectangular Box

针对海岸、岛礁、码头等因素干扰而造成的SAR图像近岸舰船检测精度不高问题,设计了采用旋转矩形框与卷积注意力模块CBAM(Convolutional Block Attention Module)改进RetinaNet的SAR图像近岸舰船检测方案。该方案在RetinaNet算法基础上,利用具有目标角度参数的旋转矩形框减弱非目标区域对舰船特征提取的干扰,在RetinaNet特征提取网络相邻残差块之间加入卷积注意力模块进行目标特征的有效聚焦,从而改善近岸舰船检测效果。利用公开的SSDD数据集、自标注近岸数据集进行了舰船检测实验,得到的检测精度相较于常规RetinaNet算法分别提升了7.02%和8.89%,验证了该方案的有效性。

Aimed at the problem that the detection accuracy of SAR images nearshore ship is not high due to the interference of factors such as coasts,reefs and wharfs,a scheme of SAR images ship detection that uses rotated rectangular box to reduce the background interference and adds Convolutional Block Attention Module to focus the target is proposed in the paper.Based on the RetinaNet framework,a rotated rectangular box with target angle parameter is used to reduce the interference of non-target area on the feature extraction of the ship,and a convolutional block attention module is added in to adjacent residual blocks of the RetinaNet feature extraction network to focus the target features effectively.The SSDD dataset and self-annotated nearshore dataset are used to conduct the ship target detection experiment.Compared with RetinaNet algorithm,the detection accuracy obtained is improved by 7.02%and 8.89%respectively,which verifies the effectiveness of the proposed scheme.