A 36-filter Hubble star-field benchmark for super-resolution: 35.42 dB PSNR on unseen filters and flux error down to 0.4341
Synopsis
Astronomical super-resolution turns low-resolution observations into high-resolution star fields, but earlier benchmarks covered only a few filters and rarely tested whether a model keeps stellar brightness on filters it never saw in training. STAR++ builds about 1.7 million flux-consistent low-resolution/high-resolution image pairs from 36 Hubble Space Telescope filters and adds a flux-error metric; FISR++, trained on it, reaches 35.42 dB PSNR and 0.4341 flux error on the unseen-filter test set, 1.1 dB higher PSNR and 0.0963 lower flux error than FISR under the same setting.
Interpretation
Astronomical super-resolution now has a large-scale benchmark spanning 36 Hubble filters and about 1.7 million flux-consistent image pairs, and it treats whether reconstructed stars keep their total brightness as a separate evaluation metric. Earlier astronomical super-resolution datasets covered at most a few filters; STAR held 54,738 pairs from a single filter and lacked systematic evaluation on filters absent from training. STAR++ generates 1,697,816 low-resolution/high-resolution pairs from 2,019 Hubble science images, with pivot wavelengths from 222.4 to 906.0 nm, and reserves seven filters absent from training as interpolation and extrapolation test sets.
The companion model FISR++ improves both reconstruction quality and flux consistency on unseen filters, beating FISR under the same training setting and general super-resolution methods including EDSR, RCAN, SwinIR and HAT. FISR trained on STAR++ reaches 34.32 dB overall PSNR and 0.5304 flux error; FISR++ raises overall PSNR to 35.42 dB and lowers flux error to 0.4341. Evaluation runs on the unseen-filter test sets of STAR++, where the interpolation subset reaches 38.05 dB PSNR and the extrapolation subset 31.91 dB, 0.83 dB and 1.45 dB above FISR.
The model handles varying spectral conditions and the strong spatial ambiguity of stellar point sources by injecting filter physical parameters into image features and separating sub-pixel position uncertainty from the reconstruction. Earlier astronomical super-resolution networks took mainly image input without explicitly using filter properties such as wavelength and bandwidth, and did not explicitly model point-source position ambiguity. Ablation shows the filter-aware adapter alone raises PSNR by 0.32 dB and lowers flux error from 0.5304 to 0.5207; adding the filter-aware adapter together with the full spatial ambiguity disentangle loss raises PSNR by 0.95 dB over the baseline and lowers flux error to 0.4536.
Perspective
The evaluated scope is science images from the Hubble Space Telescope's ACS/WFC and WFC3/UVIS imagers; low-resolution inputs are synthesized by point-spread-function blurring followed by flux-conserving downsampling, and the test sets are seven filters absent from training, split into interpolation and extrapolation groups.
The low-resolution side comes from simulated degradation of high-resolution Hubble images rather than real low-resolution telescope observations; the global pixel shift in the spatial ambiguity disentangle loss is estimated only during training and is not applied to the reconstructed image at inference.
