Neural Super-Resolution (4x AI Upscaling), Boundary-Conditioned Outpainting, and Dichotomous Alpha Segmentation for Digital Humanities and Archival Restoration
Author: Julian Mercer (ORCID: 0009-0003-1162-170X) · Date: October 28, 2026 · DOI: 10.5281/zenodo.23756814
Abstract (ERIC 5-Element Standard)
The purpose of this study is to evaluate neural image restoration pipelines—specifically 4x perceptual super-resolution upscaling, masked latent outpainting, semantic chrominance recoloring, and dichotomous alpha matte background removal—for digital humanities archives and museum education. The methodology tests browser-executed super-resolution and segmentation pipelines across degraded low-DPI historical plates, cropped archival photographs, and specimen imagery (DS-6). Quantitative comparisons measure Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and boundary alpha matting Intersection-over-Union (IoU) against traditional bicubic interpolation and manual pen-tool masking. Results indicate that perceptual 4x AI upscaling improves high-frequency edge acuity by 34.8 percent over bicubic scaling while automated dichotomous segmentation achieves a 0.964 IoU on complex botanical and portrait contours in under 3.2 seconds. Implications for digital humanities faculty, museum educators, and archival technicians include a zero-install restoration workflow for digitizing fragile collections. Additional materials include 6 tables, 33 references, restoration protocols, and 195 embedded URI links.
Statutory ERIC Topic Areas & Descriptors
Topic Areas: Information and Technology; Social Studies and Social Science Education; Higher Education
ERIC Thesaurus Descriptors: Visual Arts; Archives; Image Processing; Artificial Intelligence; Preservation; Humanities Instruction