BROWSE
Detail
Conference Paper
한국어 장면 텍스트 인식에서 VLM의 시각적 역량 및 강건성 분석
Cited - time in 
Share 
Share
- Issue Date
- 2026-04
- Citation
- 지능정보 및 제어 학술대회 (IICC) 2026, pp.42-43
- Publisher
- 대한전자공학회/대한전기학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Scene Text Recognition (STR) aims to recognize text within images captured in natural environments and serves as a fundamental component for various downstream applications. While the rapid advancement of Large Vision-Language Models (VLMs) has enabled unified visual understanding and text recognition, systematic analysis of VLM-based STR performance for non-Latin scripts, particularly Korean, remains under-explored. In this study, we evaluate the visual competency and robustness of various VLMs by establishing a Korean scene text benchmark dataset featuring diverse visual perturbations. Our experiments assess model performance under several degradations, including blur, occlusion, rotation, and perspective distortion. The results demonstrate that general-purpose VLMs generally exhibit improved robustness to visual variations compared to traditional OCR approaches, while also revealing model-specific error patterns.