Super-Resolution Analysis of Remote Sensing Images Based on Cross-Attention
International Journal of Multimedia Computing · 2024
To achieve the goal of enhancing the resolution of Sentinel-2 satellite remote sensing images by a factor of four, this paper innovatively proposes a super-resolution model (S2SR) that combines VMamba and Transformer technologies.By skillfully introducing the Mixed Attention Block (MTB) and Cross Attention Block (CAMB), the model effectively integrates the channel attention mechanism with two-dimensional selective scanning technology.This design not only enhances the synergistic utilization of global and local features but also significantly improves the interaction capability of cross-window information through the overlapping cross-attention mechanism, effectively suppressing the common block effect issue in traditional super-resolution methods and thereby significantly enhancing the quality of reconstructed images.Experimental results demonstrate that on the SEN12MS standard dataset, the S2SR model exhibits superior performance compared to existing advanced methods in multiple no-reference image quality assessment metrics (such as NIQE, BRISQUE, PIQE).Especially when processing images with complex geographical features, the super-resolution images generated by the S2SR model exhibit clear edges and rich details, fully verifying the efficiency and practicality of the model.