Scholarlink

<p>Text-based visual question answering (TextVQA) extends conventional visual question answering (VQA) by requiring integrated reasoning over visual features, scene text, and natural language questions. Despite substantial progress in recent years, challenges remain in robust text recognition, effective multimodal fusion, spatial alignment, and computational efficiency. This paper presents a systematic review of TextVQA methodologies and categorizes existing approaches into four classes: attention-based, transformer-based, graph-based, and frozen large language model (LLM)-based frameworks. Representative models from each category are analyzed in terms of architectural design, text integration strategies, and reasoning mechanisms. A comparative study is conducted across widely used benchmark datasets, including TextVQA, Scene Text Visual Question Answering (ST-VQA), and Optical Character Recognition Visual Question Answering (OCR-VQA), using standard evaluation metrics. The review highlights the methodological evolution of the field, illustrating the transition from early optical character recognition (OCR)-dependent attention architectures to transformer- and graph-based models with enhanced cross-modal interactions, and more recently to LLM-oriented systems that emphasize scalability and improved multimodal reasoning. Finally, key research challenges are identified, and future research directions are outlined to guide the development of more robust, scalable, and efficient TextVQA systems.</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

The evolution and open challenges of text-based visual question answering: a review of research and data trends

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Kobra Farshidi,
  • Hassan Khotanlou,
  • Elham Alighardash

摘要

Text-based visual question answering (TextVQA) extends conventional visual question answering (VQA) by requiring integrated reasoning over visual features, scene text, and natural language questions. Despite substantial progress in recent years, challenges remain in robust text recognition, effective multimodal fusion, spatial alignment, and computational efficiency. This paper presents a systematic review of TextVQA methodologies and categorizes existing approaches into four classes: attention-based, transformer-based, graph-based, and frozen large language model (LLM)-based frameworks. Representative models from each category are analyzed in terms of architectural design, text integration strategies, and reasoning mechanisms. A comparative study is conducted across widely used benchmark datasets, including TextVQA, Scene Text Visual Question Answering (ST-VQA), and Optical Character Recognition Visual Question Answering (OCR-VQA), using standard evaluation metrics. The review highlights the methodological evolution of the field, illustrating the transition from early optical character recognition (OCR)-dependent attention architectures to transformer- and graph-based models with enhanced cross-modal interactions, and more recently to LLM-oriented systems that emphasize scalability and improved multimodal reasoning. Finally, key research challenges are identified, and future research directions are outlined to guide the development of more robust, scalable, and efficient TextVQA systems.