Uncertainty‐Aware Processing for OCR Using Dictionary Routing and Candidate Selection

ABSTRACT

Optical Character Recognition (OCR) systems often produce errors due to ambiguous handwriting and character similarity. This paper proposes an Uncertainty‐Aware Processing (UAP) framework that reduces the Character Error Rate (CER) without additional training or inference. The method calculates and metrics using the results of connectionist temporal classification (CTC). These two metrics are used to estimate a character‐level uncertainty. The UAP framework selectively applies correction by reusing CTC outputs. Experiments on three datasets (CVL, IAM and GW) show CER improvements across three CTC‐based models. In particular, SVTRv2_mobile model demonstrated the effectiveness of the proposed method by reducing the CER from 17.08% to 12.67% on the IAM dataset.