An Efficient Reference-Guided Soiled Region Segmentation Network for Unmanned Surface Vehicles

Fast and accurate segmentation of soiled regions caused by lens obstructions is not only crucial for enhancing the autonomy of unmanned surface vehicles (USVs) but also vital for ensuring safety. However, most existing semantic segmentation methods are proposed for natural scene segmentation and are difficult to be directly applied to the detection of soiled regions with various shapes and blurred edges. In this paper, an efficient reference-guided soiled region segmentation network (SRSNet) for USVs is proposed, which improves the segmentation of soiled regions by integrating prior information from the reference image of adjacent viewpoints. Specifically, we devise a novel spatial alignment and attention module (SAAM) that employs a deformable convolution to model spatial correspondences between the reference and soiled images, thereby aligning their features to enhance soiled region representation learning. To effectively aggregate multi-scale aligned features from different stages, a cross-stage feature fusion module (CFFM) based on channel attention is presented. Both quantitative and qualitative evaluations demonstrate that the proposed SRSNet outperforms state-of-the-art methods in soiled region segmentation while running in real-time.