Small targets recognition in SAR ship image based on improved SSD

Abstract

Synthetic Aperture Radar(SAR) ship recognition is significant in marine applications and plays an important role in maritime traffic management, fisheries management, and maritime rescue, etc. A major difficulty in SAR ship recognition is that the SAR ships have a small size in images, which results in low recognition accuracy of the SSD. This paper first analyzes the reasons for the low recognition accuracy for small targets in SSD. First, the poor matching of the default boxes leads to a small number of positive samples. Second, the representation <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</sup> bility of low-level feature map for recognizing small targets is weak. Then two strategies are proposed to improve the SSD. Firstly, a default box optimization design method based on Kmeans clustering is proposed, which improves the matching performance of the default boxes. Secondly, a feature fusion method based on deconvolution is proposed, which effectively improves the representation ability of low-level feature maps. The experimental results show that the proposed method can greatly improve the recognition accuracy of SSD for small targets in SAR ship images.