Answers are grounded in this paper and cite their sources.
Remote sensing image super-resolution requires local detail and wider structural context. The mixture-of-experts (MoE) designs compared here select or combine expert outputs without explicitly passing behavior statistics to subsequent spatial–frequency restoration. We propose EBSFFormer, which aggregates routing preferences and response-weighted proportions within restoration groups and converts them into channel gains for later spatial and frequency branches. Heterogeneous experts and a content–cumulative-change router supply these statistics. We train on AID and evaluate on AID, DIOR, DOTA, and NWPU-RESISC45 under bicubic ×4 degradation. Across three AID training seeds, EBSFFormer obtains 31.2639 ± 0.0009 dB peak signal-to-noise ratio (PSNR), with mean improvements of 0.0121 dB over MambaIRv2 and 0.1037 dB over SwinIR. Single-checkpoint evaluation on NWPU without fine-tuning gives the highest Y-channel PSNR and structural similarity index (SSIM) among the compared methods and the third-lowest learned perceptual image patch similarity (LPIPS). Component ablations, red–green–blue (RGB) metrics, and ground-truth-referenced visual comparisons characterize the reconstruction gains. EBSFFormer has fewer parameters and lower mean latency than MambaIRv2, but the highest peak allocated memory in the runtime comparison.
No takes yet. Share an insight, caveat, or question.
Xie et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: