PageNet: Towards End-to-End Weakly Supervised Page-Level Handwritten Chinese Text Recognition

Handwritten Chinese text recognition (HCTR) has been an active research topic for decades. However, most previous studies solely focus on the of cropped line images, ignoring error caused by detection in real-world applications. Although some approaches aimed at page-level have proposed recent years, they either are limited to simple layouts or require very detailed annotations including expensive line-level and even character-level bounding boxes. To this end, we propose PageNet end-to-end weakly supervised HCTR. detects recognizes characters predicts reading order between them, which is more robust flexible when dealing with complex multi-directional curved lines. Utilizing learning framework, requires only transcripts be annotated real data; however, it can still output results both character levels, avoiding labor cost labeling boxes Extensive experiments conducted five datasets demonstrate superiority over existing fully methods. These experimental may spark further beyond realms methods based connectionist temporal classification attention. The source code available https://github.com/shannanyinxiang/PageNet .