TextVQA : A Comprehensive Study of Approaches, Datasets, and Evaluation Metrics

Text‐based visual question answering (Text‐VQA) aims to answer question by combining visual features with embedded textual content. This field has gained substantial attention due to developments in deep learning technologies across vision, text and voice. A typical Text‐VQA pipeline includes Optical Character Recognition (OCR) for extracting textual content from images, which are then employed within a VQA model to reason about the content. In this survey, we conducted a detailed study of the field, systematically analyzing key approaches, datasets, and evaluation metrics that have shaped the expansion of Text‐VQA systems. We provide a comprehensive overview of recent state‐of‐the‐art (SOTA) methods and evaluate their performance on three major datasets: TextVQA, ST‐VQA and OCR‐VQA. By observing each phase of the Text‐VQA methods, we discussed key challenges, acknowledge major contributions, and presented promising directions for future research in this evolving field.