Uncertainty-Aware structured decoding for Chinese Braille document transcription

Automatic transcription of braille documents remains challenging in document digitization settings, particularly under real imaging degradations such as uneven embossing, illumination variation, paper texture, skew, blur, and ghosting, which make braille dot presence uncertain and often cause error propagation from dot detection to cell encoding and final text transcription. To address this problem, we propose an uncertainty-aware structured decoding framework for Chinese braille document transcription that formulates the task as document analysis with an explicit intermediate representation rather than as a purely end-to-end black-box mapping. The method first performs robust grid fitting and structured dot analysis within braille candidate regions, and converts dot responses into calibrated probabilities. Topological consistency constraints are then imposed to recover valid six-dot cell patterns while producing dot-level and cell-level confidence estimates for risk localization. To resolve residual ambiguity efficiently, only low-confidence cells are expanded into small local candidate sets, which are rescored using contextual compatibility and local structural cost. Experiments on authentic scanned and photographed braille documents using document-level five-fold cross-validation achieved a Dot-F1 of 0.992 ± 0.002, a character error rate of 0.018 ± 0.003, and a BLEU-4 score of 74.1 ± 1.2. The results demonstrate that the proposed framework improves structural recovery stability, transcription accuracy, and sentence-level consistency, while maintaining strong robustness across document conditions and limiting additional inference overhead.