DiffSARShipInst: Diffusion model for ship instance segmentation from synthetic aperture radar imagery

Abstract

Recently, deep learning (DL) methods, particularly convolutional neural networks (CNNs)-based ones, have significantly advanced the development of synthetic aperture radar (SAR) ship instance segmentation. However, existing instance segmentation algorithms typically rely on preset candidate boxes, which are challenging to perfectly match to ships from a regression optimization perspective, limiting segmentation accuracy. Therefore, we propose a novel diffusion model, DiffSARShipInst, for SAR ship instance segmentation. This model represents ship instance segmentation as a denoising process from noisy boxes to target boxes and a reconstruction process from target boxes to ship instances. It innovatively handles the ship instance segmentation task from a generative perspective, treating random boxes as object candidates to overcome the limitations of existing methods that require target priors. To achieve superior SAR ship instance segmentation accuracy, DiffSARShipInst also offers: 1) a spatial-contextual joint enhanced feature pyramid network (SCJE-FPN) to improve the multi-scale ship feature extraction ability for the subsequent denoising and reconstruction processes; 2) a focused intersection-over-union (FIoU) loss to suppress redundant noisy samples during the denoising process; and 3) an instance-aware mask representation (IAMR) to adaptively generate ship instances from denoised target boxes during the reconstruction process. Extensive experiments on the SAR ship detection dataset (SSDD) and the high-resolution SAR image dataset (HRSID) demonstrate its superior performance. Specifically, DiffSARShipInst achieves up to 70.6 %/70.9 % mask average precision (AP) in offshore scenes of SSDD/HRSID, and 56.2 %/42.6 % mask AP in inshore scenes of SSDD/HRSID.