SAR Ship Instance Segmentation Based on Hybrid Task Cascade

Abstract

Ship detection in synthetic aperture radar (SAR) images has received extensive attention in recent years. Yet, SAR ship instance segmentation is rarely studied. Therefore, we apply the hybrid task cascade (HTC) delving into the targeted ship in-stance segmentation from SAR imagery. HTC comprehensively considers the coupling relationship between the detection task and segmentation task. Specifically, it replaces the original single detection-head (DH) of the raw Mask R-CNN with three cascaded DHs to further improve performance. We perform many experiments based on the public SAR Ship Detection Dataset (SSDD) dataset for verifying the effectiveness of HTC. Finally, it offers a 65.6% detection $AP$ that is superior to Mask R-CNN by 3.4%, and provides a 59.3% segmentation $AP$ that is superior to Mask R-CNN by 1.5% as well.