Answers are grounded in this paper and cite their sources.
Post-OCR correction of Chinese historical books is bounded by substitution errors, yet the glyph cues that would resolve them are used poorly: hand-crafted descriptors break down across the domain gap between woodblock printing and modern typefaces, and they are keyed to the recognized character rather than to the observed glyph image. LGE-FIng learns a glyph embedding under an ArcFace objective with cross-domain alignment, so that a woodblock glyph block and a multi-font rendered prototype of the same character share one class center. At each position the embedding is queried by the real image block, and its nearest neighbors feed image-conditioned candidates into the corrector's search lattice. On a 60-page book-disjoint test set, the image-queried channel beats an identical channel keyed to the recognition character's rendering by + 1.13 to + 1.54 percentage points (pp) under four independent weight-mapping calibrations, all agreeing in sign and magnitude. The gain survives a scale-free rank calibration that equalizes the two channels' weight ladders, which rules out a cosine-scale artifact. A layout-aware fusion of both channels reaches + 3.66 pp over the bare recognizer (PP-OCRv5) and + 3.73 pp with a real-corpus hybrid gallery. A mechanistic probe traces the effect to rank one: the image channel ranks the ground truth first in 19.4% of substitution positions, while the static channel cannot do so by construction. The method trains cheaply, stays OCR-agnostic, and integrates into an existing corrector as an additional candidate source.
No takes yet. Share an insight, caveat, or question.
shen et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: