A Pixel-Level Segmentation Method for Water Surface Reflection Detection

Water surface reflections pose challenges to unmanned surface vehicles or robots during target detection and tracking tasks, leading to issues such as the loss of tracked targets and false target detection. Current methods for water surface reflection detection primarily rely on image thresholding, saturation, edge detection techniques, which have poor performance in segmentation, as they are more suitable for handling simper image scenarios and are insufficient for the detection of water surface images characterized by complex background information, intricate edge details, and the inclusion of abundant contextual elements from both shores. To bridge the gap, we propose a novel model named WRS-Net for achieving pixel-wise water reflection segmentation, which leverages an encoder-decoder architecture and incorporates two novel modules, namely Multi-scale Fusion Attention Module (MSA) and Interactive Convergence Attention Module (ICA). In addition, a water surface reflection dataset for sematic segmentation is constructed. The MSA extracts detailed local reflection features from shallow networks at various resolutions. These features are subsequently fused with high-level semantic information captured by deeper networks, effectively reducing feature loss and enhancing comprehensive extraction of both shallow features and high-level semantic information. Additionally, the ICA consolidates the preservation of local reflection details while simultaneously considering the global distribution of the reflected elements, by encapsulating the outputs of the MSA, the multiple feature maps of various scales, with the outputs of the decoder. The experiment results demonstrate enhanced performance of the proposed method in contour feature extraction and effective reflection segmentation capabilities. Specifically, the proposed method achieves mIoU, mPA, and average accuracy of 94.60%, 97.70%, and 97.96%, respectively, on the water reflection semantic segmentation dataset.