Oriented Bounding Box Representation Based on Continuous Encoding in Oriented SAR Ship Detection

Abstract

Ship detection of synthetic aperture radar (SAR) images holds significant value for both civilian and military applications. Compared to horizontal ship detection, oriented ship detection based on oriented bounding boxes can capture the orientation and aspect ratio of ships, thus receiving increasing attention. However, in oriented ship detection, when ships rotate near the specific angles, the angle prediction result obtained by the deep learning network may have a severe mutation, which is the well-known boundary discontinuity problem. To address the issue of boundary discontinuity in SAR ship detection, researchers have proposed numerous methods. However, through our systematic analysis, we found that these methods do not fundamentally solve the problem. To this end, we first clarified the reasons for the existence of boundary discontinuity and how it affects the detection network. Based on this, we proposed the conditions that the encoding methods and loss functions of the detection network must satisfy to address the issue of boundary discontinuity. In line with these conditions, we designed a continuous encoding method called coordinate decomposition method (CDM). In addition, we also analyzed the impact of different optimization methods on the detection network and, based on this, presented a joint optimization paradigm based on continuous encoding. Experimental results on two commonly used SAR ship detection datasets demonstrate that our proposed CDM encoding method effectively addresses the boundary discontinuity issue and enhances the detection performance. Compared to the state-of-the-art methods, the fully convolutional one-stage network using the CDM-based joint optimization achieves optimal detection results without employing any additional techniques.