DSCT: Dual-Sparse Coherent Transformer With Cross-Feature Alignment for Remote Sensing Image Super-Resolution

Le résumé fourni par la source

Transformer-based methods have achieved significant progress in remote sensing image super-resolution (SR), benefiting from the flexible formulation of self-attention mechanisms for modeling long-range dependencies and structural correlations. However, their performance is still constrained by two key issues. First, redundant spatial and channel responses may introduce irrelevant activations during attention computation, diluting informative regions and hindering effective feature aggregation, thereby degrading fine-grained structure recovery. Second, existing approaches often rely on coarse feature fusion strategies, which fail to explicitly model the coherence and complementary relationships between heterogeneous feature dimensions, leading to suboptimal representations. To address these limitations, we propose the dual sparse coherent Transformer (DSCT), which jointly models adaptive sparsity and feature coherence for more effective representation learning. Specifically, DSCT adopts an alternating design that integrates the adaptive sparse spatial self-attention (ASS-SA) and the selective sparse channel self-attention (SSC-SA) to enable complementary modeling across spatial and channel dimensions. ASS-SA enhances dominant spatial correlations while suppressing redundant responses by adaptively reweighting attention distributions under different sparsity patterns, whereas SSC-SA dynamically regulates interactions across channels through sparsity modulation guided by input-aware cues. In addition, a cross-feature alignment (CFA) mechanism and a feature modulation feedforward network (FMFN) are introduced to enhance feature coherence and representation quality, where CFA calibrates and aligns complementary features before fusion, and FMFN injects selective spatial modulation together with gated feature interactions for adaptive feature refinement. Extensive experiments on multiple remote sensing SR benchmarks demonstrate that DSCT achieves superior performance compared with state-of-the-art Transformer-based methods while maintaining competitive efficiency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.