SW 중심대학 공동 AI 경진대회 (본선) - Optical Character Recoginition (OCR)

• 팀 경하예프 (경희대)
• 참가자: 김민성, 박수용, 이상민, 차준영, 한주혁
• Private Score (Accuracy): 92.44% (최종 10등)
• 발표 평가 및 최종 순위: 5등

Overview

광학 문자 인식 Task로 주어진 데이터를 보고 어떤 문제가 있고, OCR 모델의 정확도를 향상시키기 위해선 이 문제들을 어떻게 해결해야 하는지 위주로 접근했습니다.

Reference

• TPS-Resnet-BiLSTM-Attn: BAEK, Jeonghun, et al. What is wrong with scene text recognition model comparisons? dataset and model analysis. In: Proceedings of the IEEE/CVF international conference on computer vision. 2019. p. 4715-4723.

(Official code: https://github.com/clovaai/deep-text-recognition-benchmark)

• SATRN: LEE, Junyeop, et al. On recognizing texts of arbitrary shapes with 2D self-attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020. p. 546-547.

(Not official code, but we applied this code to train OCR model: https://github.com/Media-Smart/vedastr)

Library

....

Problem Solving Strategy