Enhancing ship wake segmentation via attention-guided feature refinement and spatial interaction

Wake separation for ships is important in both civilian and military applications. In recent years, image segmentation techniques are widely applied in the extraction and analysis of the wake in optical remote sensing images. The wake is usually characterized by features such as elongated morphology and corrugated structure. While performing wake segmentation, challenges such as poor global modelling results, loss of texture detail information and blurred edges are commonly encountered. To address the above problems, this paper proposes an attention-guided feature refinement and spatial interaction network (FRSINet) and specifically designs and incorporates three tailored enhancement modules. First, we fuse K-Net with several classic semantic segmentation models. This builds a hybrid model architecture which is most suitable for the ship wake segmentation task. Subsequently, we introduce convolutional block attention module-neck (CBAM-N) and feature refinement and enhancement module (FREM) into the neck network of the hybrid model. Finally, the spatial interaction module-residual (SIM-R) is incorporated into the backbone network. The experimental results on the SWIM-S dataset demonstrate that, FRSINet achieves more accurate ship wake segmentation than other methods.