한글 자모 기반 OCR을 활용한 송장 문자 인식

초록

As the automation of the logistics industry accelerates due to recent advancements in robotics and AI technology, the need for accurate character recognition in invoice processing has become increasingly critical. However, conventional OCR, which relies on standardized formats, struggles to handle unstructured invoices. To overcome these limitations, this study proposes a sub-character-based OCR model. The proposed model utilizes a pretrained ViT as the encoder and an 8-layer transformer architecture for the decoder. Additionally, an Additive Gated Bridge is implemented between the encoder and decoder to effectively transmit meaningful features. The training process consists of pretraining based on synthetic data followed by fine-tuning on actual invoice data, with samples failed during recognition reintegrated into the retraining process for each epoch. Consequently, the proposed method achieved an exact match accuracy of 0.92 on the evaluation dataset.

키워드