Handwritten Text Recognition: A Comprehensive Survey of Evolution and Architectures

Abstract

This survey presents a comprehensive synthesis of contemporary research in handwritten text recognition (HTR), encompassing bibliometric analysis, systematic methodological reviews, and novel architectural innovations spanning line-level, document-level, and multi-lingual recognition systems. Drawing from 25 original research works, survey papers, and empirical studies, this work categories existing literature into ten thematic clusters: survey and review studies, meta-learning and adaptation methods, self-supervised learning approaches, vision-language models, transformerbased architectures, word and keyword spotting methods, document-level recognition systems, low-resource and Indic script recognition, foundational deep learning architectures, and out-of-distribution generalization studies. The survey reveals that while significant progress has been achieved through deep learning and transfer learning techniques, critical challenges persist in handling domain shifts, low-resource languages, and complex document layouts. Furthermore, emerging paradigms including self-supervised vision transformers, meta-learning frameworks, and foundation models demonstrate substantial promise for enabling more adaptive, efficient, and generalisable HTR systems. This paper synthesises these developments, identifies cross-cutting methodological themes, analyses comparative performance trends, and delineates key research gaps that warrant future investigation.