We recovered 575k crop labels from a decade of manual Photoshop work to automate book digitization - more data, ResNet-50, and higher resolution all failed; ten operator clicks per book beat them [P]

Author here. Ibteda Digital Library is a private community archive in Pakistan — for ten years we digitized rare Urdu books (lithographs, dictionaries, periodicals) on a DIY camera rig, finishing every page by hand in Photoshop. When we wound down daily operations, I realized those 575,729 finished pages across 1,765 books recorded a decade of crop decisions, so I registered them back to their raw photos (SIFT + MAGSAC with conservative acceptance gates) and used the recovered geometry as supervision. The negative results are probably the most interesting part for this sub. Scaling from 378 to 572 training books didn't move unseen-book pass@80. Neither did ResNet-50 (better training fit, flat held-out, worse after calibration), 1024px inputs, or a spatial head. Per-book error analysis showed why: the failures were near-constant offsets per volume — our operator's preferred margin inset, which simply isn't present in the pixels of a new book. Ten operator-corrected crops per book (element-wise median residual) took pass@80 from 0.71 to 0.83 on held-out volumes. Ten labels beat every scaling lever we tried. For retouching (stain/stamp removal), we kept the neural net to detection only — a U-Net proposes removal support, classical OpenCV reconstructs the paper, and everything outside the mask is byte-identical to the original. Labels used REMOVE/KEEP/IGNORE states, and any erased Urdu diacritic vetoed deployment regardless of IoU. The stricter label cut both improved mark IoU (0.56 → 0.60) and got diacritic false positives to zero. Two things I'd genuinely like input on: (1) has anyone modeled document boundaries that depend on an invisible human preference rather than visible structure — is there prior work on per-instance residual calibration like this? Our own next step is conditioning the model on the calibration examples directly (few-shot inset inference) instead of a post-hoc median. (2) Is there any constrained diffusion/inpainting setup you'd trust to guarantee zero alteration outside a declared support region, or is classical fill still the only honest option for archival work? Full training recipes, label-mining thresholds, and routing rules are in the Reproduction Contract appendix. Code/weights release is still under archival review — I know, I'm sorry — but I'm happy to answer anything about the method here. Notably not on our roadmap: a bigger backbone. We tested that; the missing information isn't in the pixels. Link