CEASeg: Cross-modal Efficient Attention Semantic Segmentation Method for Remote Sensing Images

Cross-modal semantic segmentation plays a significant role in Remote Sensing due to the complementary and heterogeneous nature of multi-source data. Nevertheless, existing cross-modal segmentation methods tend to extract shared features with complex network structures, which restrict the efficiency of segmentation and lead to considerable computational redundancy. To overcome the limitations, a cross-modal efficient attention segment model (CEASeg) for RS images is designed. CEASeg first extracts multi-level features through a well-designed dual-branch network. Next, we design two efficient blocks for cross-modal fusion. The attention-based cross-modal fusion block (ACMFusion) can learn feature correlations between two modalities and perform weight up-sampling, while the partial cross-attention fusion block (PCAFusion) emphasizes the global interaction information of deep features. The promising experimental results on two datasets show that CEASeg outperforms the state of the art approaches in both accuracy and efficiency.