Scholarlink

<p>Scene text recognition (STR) in natural images remains challenging due to variations in orientation, curvature, illumination, and background complexity. Conventional Optical Character Recognition (OCR) systems, which assume clean, well-structured text, struggle to generalize to such real-world conditions. This work presents a transformer-based, two-phase framework for robust scene text extraction that addresses geometric distortions and arbitrary text orientations. The proposed method integrates a ResNet50-based detection module with a Vision Transformer (ViT) recognition model combined with a Connectionist Temporal Classification (CTC) decoder, enabling accurate word-level prediction without relying on character-level annotations. In the first phase, text regions are localized using an enhanced contour-based detection mechanism that incorporates boundary refinement and a Local Context Map (LCM) to improve discrimination between text and background clutter. In the second phase, detected regions are processed by a ViT encoder that models long-range dependencies, followed by CTC decoding to generate variable-length text sequences efficiently. The framework is evaluated on multiple public benchmarks, including ICDAR15 and SCUT-CTW1500, demonstrating strong robustness to rotation, curvature, and complex backgrounds. Comparative experiments show competitive recognition performance and notably high recall in detection, despite using only word-level supervision. The findings underscore the effectiveness of transformer architectures in unconstrained STR and highlight the potential of the proposed system for real-time applications such as signage interpretation, autonomous navigation, and assistive technologies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Rotation-invariant scene text extraction using transformer networks

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Vanitha Sivagami Sivasankaravel,
  • Sreenivasan S R,
  • Manoj R

摘要

Scene text recognition (STR) in natural images remains challenging due to variations in orientation, curvature, illumination, and background complexity. Conventional Optical Character Recognition (OCR) systems, which assume clean, well-structured text, struggle to generalize to such real-world conditions. This work presents a transformer-based, two-phase framework for robust scene text extraction that addresses geometric distortions and arbitrary text orientations. The proposed method integrates a ResNet50-based detection module with a Vision Transformer (ViT) recognition model combined with a Connectionist Temporal Classification (CTC) decoder, enabling accurate word-level prediction without relying on character-level annotations. In the first phase, text regions are localized using an enhanced contour-based detection mechanism that incorporates boundary refinement and a Local Context Map (LCM) to improve discrimination between text and background clutter. In the second phase, detected regions are processed by a ViT encoder that models long-range dependencies, followed by CTC decoding to generate variable-length text sequences efficiently. The framework is evaluated on multiple public benchmarks, including ICDAR15 and SCUT-CTW1500, demonstrating strong robustness to rotation, curvature, and complex backgrounds. Comparative experiments show competitive recognition performance and notably high recall in detection, despite using only word-level supervision. The findings underscore the effectiveness of transformer architectures in unconstrained STR and highlight the potential of the proposed system for real-time applications such as signage interpretation, autonomous navigation, and assistive technologies.