SAR Ship Detection Based on An Improved CNN with IoU-FL and Attention Mechanism

Abstract

Ship detection in synthetic aperture radar (SAR) images is important in the area of marine surveillance. However, the extremely unbalanced distribution of easy and hard examples leads to difficulty in accurately profiling and locating ships in SAR images. In this paper, we propose an improved convolutional neural network (CNN) based on a novel intersection over union (IoU)-embedded-focal loss (IoU-FL) and a convolutional block attention module (CBAM). To reduce the impact of easy background examples and focus on learning hard examples, the focal loss is introduced in the proposed method, during the training process. During the regression process, a novel IoU-embedded- focal loss (IoU-FL) is proposed to mitigate the imbalance problem in bounding box regression (BBR). Moreover, three CBAM modules are embedded in the backbone of the network to refine different level semantic feature maps and suppress unnecessary background ones by focusing on ‘what’ is meaningful in the channel and ‘where’ is informative in the spatial. The proposed method is evaluated on the public SAR ship detection dataset (SSDD) and the results demonstrate that the proposed method is superior to conventional methods in efficiency and accuracy.