Towards Robust Handwritten Chinese Text Understanding: Benchmarks and Unified Frameworks for Detection, Recognition, and Restoration
Year: 2025
Type: dissertation
Abstract: Handwritten text appears widely in real-world scenarios such as student notes, text book pages, packaging bags, cardboard boxes, and historical documents. It often presents challenges, including missing characters, background noise, and complex layouts. Additionally, variations caused by image acquisition, such as multi-angle shooting, uneven exposure, and low contrast, further reduce the legibility and hinder the digitization and automatic understanding of handwritten content. Most existing methods focus on line-level recognition under relatively clean backgrounds, while overlooking the importance of glyph information, which is also essential in fields such as art, culture, and historical studies. In particular, current approaches lack ef fective modeling for glyph restoration in complex environments. To address these challenges, this work emphasizes glyph restoration alongside text recognition. A benchmark dataset of handwritten Chinese text is constructed, covering a wide range of noisy scenarios with synthetic and real-world samples. It supports line-level and page-level granularity and includes a glyph annotation tool with high-quality annotations for related research. Methodologically, a base line model named ResLU is proposed for line-level text, jointly optimizing recog nition and glyph restoration via multi-task learning with the combined loss. This improves recognition robustness under noise while enabling glyph reconstruction. To overcome the limitations of line-level methods in handling layout and granular ity, a character-level, page-wise framework named UniText is further introduced. It integrates detection, recognition, and glyph restoration, remains efficient across di verse layouts and noise, and aligns well with the character-as-class nature of Chinese script. Extensive experiments on synthetic and real-world handwritten text datasets show that the proposed models achieve comparable performance to existing state of-the-art methods in text detection and recognition, while demonstrating superior effectiveness in glyph restoration, indicating good practicality and scalability.
Author: Shen, Lu
Institutions: Macao Polytechnic University, University of Coimbra
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7166995368
