Answers are grounded in this paper and cite their sources.
In feature-dense remote sensing images, super-resolution reconstruction performance relies on the ability to filter key information from complex scenes. Transformer-based methods often focus on semantically irrelevant regions due to their global self-attention mechanisms, leading to structural noise and artifacts. Although sparse attention can suppress redundant noise, its fixed threshold strategy frequently causes texture blurring by removing high-frequency details. To address the contradiction between sparsity and information retention, this article proposes the implicit state-space transformer (ISST), the first bootstrapping-optimized framework that guides sparse attention coding using semantic priors and leverages control theory inspired closed-loop state updates. Specifically, we design the implicit state-space attention block, which provides visual prompts for self-attention through the dynamic semantic correlations of implicit states. This preserves critical high-frequency information during sparse computation and avoids loss of detail in traditional methods. We further develop a segment anything model-based semantic prior prompting mechanism that generates semantic information without additional data or manual annotations. This mechanism initializes and updates implicit correlation matrices, which are then used to establish pixel correlations for fine-grained texture reconstruction via closed-loop bootstrapping feedback. Experiments demonstrate that ISST outperforms other methods on multiple datasets: in the ×4 super-resolution task on the UCMerced dataset, the average PSNR is improved by 0.60 dB compared to TTST. It effectively suppresses redundant information interference and retains details in complex scenes, maintaining high quality even at low diffusion index values.
No takes yet. Share an insight, caveat, or question.
Rong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: