Scholarlink

We present a comprehensive survey of Document Image Analysis (DIA), tracing its evolution from rule-based and feature-engineered systems through deep-learning advances to contemporary unified multimodal frameworks. The paper reviews preprocessing and segmentation fundamentals, major layout and OCR datasets (PubLayNet, DocLayNet, RanLayNet, CORD, TNCR, IIT-CDIP, D4LA), classical and deep architectures for layout and text recognition, and end-to-end vision-language models (LayoutLMv3, TrOCR, DocFormer, Donut). We discuss cross-domain degradation, trade-offs between specialized pipelines and LLM-centric approaches, and present a unified taxonomy for multimodal explainability. Finally, we propose a scalable human–AI annotation pipeline and outline research directions—cross-domain adaptivity, lightweight distillation, graph-based relational models, and hybrid human-AI systems—aimed at building accurate, efficient, and auditable DIA systems.

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

Document Image Analysis: A Review of Existing Approaches and Future Scope

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sunil Kumar Vengalil,
  • M. Rohan,
  • Khushi Nagaraj,
  • P. V. S. Sahasra,
  • Sri Likitha Nannuri

摘要

We present a comprehensive survey of Document Image Analysis (DIA), tracing its evolution from rule-based and feature-engineered systems through deep-learning advances to contemporary unified multimodal frameworks. The paper reviews preprocessing and segmentation fundamentals, major layout and OCR datasets (PubLayNet, DocLayNet, RanLayNet, CORD, TNCR, IIT-CDIP, D4LA), classical and deep architectures for layout and text recognition, and end-to-end vision-language models (LayoutLMv3, TrOCR, DocFormer, Donut). We discuss cross-domain degradation, trade-offs between specialized pipelines and LLM-centric approaches, and present a unified taxonomy for multimodal explainability. Finally, we propose a scalable human–AI annotation pipeline and outline research directions—cross-domain adaptivity, lightweight distillation, graph-based relational models, and hybrid human-AI systems—aimed at building accurate, efficient, and auditable DIA systems.