EM-UNet: Enhancing Multi-Scale Category Segmentation in Water Surface Scenes

The traditional encoder-decoder structure has been widely applied in the field of semantic segmentation. However, in tasks involving the segmentation of various water surface categories in inland rivers, the traditional encoder-decoder networks show poor robustness, lower accuracy when segmenting objects with significant scale differences, and struggle to fuse multi-scale contextual information effectively. To address these issues, we propose an improved encoder-decoder structure (EM-UNet) that enhances the ability to segment multiple categories in water-surface scenes. First, we design a multi-scale attention fusion module (EMSCA), which focuses on channel scale issues through pointwise convolution and captures pixel-level relationships through cross-spatial interaction strategies. Integrating EMSCA into the model's feature fusion part enhances the representation of features at different scales, improves the efficiency of feature fusion, and adaptively focuses on important features. Additionally, using ResN et as the backbone of the traditional structure strengthens the network's ability to capture image features and contextual information. Finally, we validate the effectiveness of EM-UNet on the WaterScenes dataset.