Degradation-parameter-supervised restoration of optical borehole images for fracture-oriented structural characterization: leakage-audited cross-domain validation

Optical borehole televiewer (OTV) images support fracture-oriented structural characterization, but non-uniform illumination can obscure weak traces. We propose PG-BIE, a 47,170-parameter restoration network that supervises illumination, transmission and airlight estimates before analytic inversion and bounded refinement. The primary dataset comprised 220 training and 55 validation rock-wall OTV images. Across five runs, PG-BIE achieved 15.146 ± 0.035 dB PSNR and 0.8384 ± 0.0044 gradient similarity under a fixed synthetic-degradation protocol. An official LOL-v1-trained Retinexformer baseline achieved 13.144 dB PSNR and 0.4627 gradient similarity and produced marked cross-domain brightening. A conservative source-group leakage audit retained 28 validation images from 21 groups; PG-BIE improved PSNR over degraded input by 2.928 dB (95% cluster-bootstrap CI, 2.081–3.621) and gradient similarity by 0.2746 (0.2346–0.3088). Independent validation used 512 windows from two ROGUE ice OPTV logs and 193 interpreted fracture bands from 11 rock borehole-television wells. The two ROGUE sites responded oppositely. On the rock bands, SSIM and gradient similarity increased, but PSNR decreased by 1.113 dB (− 1.970 to − 0.073), indicating a domain-dependent color shift. PG-BIE therefore supports traceable illumination correction, but image-quality gains do not establish improved fracture detection or orientation measurement.