Attention-Guided Digital Enhancement of Museum Artifact Photographs Using Progressive Curriculum Learning

Physical deterioration and image-quality loss can reduce the legibility of museum photographs and limit their value for digital access, cataloging, and visual analysis. This study evaluates an Attention ResUNet for grayscale enhancement under heterogeneous degradation and revisits the contribution of progressive training using compute-matched controls. A curated set of 5816 Metropolitan Museum of Art Open Access objects was divided at the object level into 4071 training, 872 validation, and 873 test objects. In the original five-epoch Phase-II ablation, staged initialization outperformed direct five-epoch training, but this advantage did not persist after total optimization exposure was matched. Direct 15-epoch ResUNet reached 29.777 dB PSNR and direct 15-epoch Attention ResUNet reached 29.862 dB, compared to 29.340 and 29.419 dB for their staged counterparts under the same 43,125-step budget. An independent set of 45 photographs captured on site at the Peshawar Museum was reserved for post-selection evaluation. Direct enhancement of the untouched photographs was assessed with no-reference indicators, with NIQE decreasing from 4.996 to 4.237 and BRISQUE from 23.171 to 18.479. A separate reference-based experiment retained each original photograph as an image-domain target, added deterministic Gaussian noise (σ=10), and restored the degraded input with the compute-matched Attention ResUNet (C15). Across all 45 photographs, mean PSNR increased from 28.149±0.045 to 29.740±0.914 dB, SSIM from 0.8136±0.0555 to 0.8985±0.0183, and LPIPS decreased from 0.1140±0.0459 to 0.0668±0.0242. SSIM and MAE improved in all 45 cases, while PSNR improved in 40/45 and LPIPS in 44/45 cases. A structured blinded assessment on 15 paired cases found the C15 output closer to the pre-corruption reference in 14/15 comparisons; a second independent reviewer provided qualitative corroboration but did not complete a second image-level score sheet. The paired experiment establishes recovery from a known image-domain corruption, not reconstruction of the artifact’s historically pristine physical state. The method remains limited to grayscale structural and luminance enhancement, and processed outputs should remain linked to the original photographs.