Recognition of Handwritten Chinese Text by Segmentation: A Segment-Annotation-Free Approach

Online and offline handwritten Chinese text recognition (HTCR) has been\nstudied for decades. Early methods adopted oversegmentation-based strategies\nbut suffered from low speed, insufficient accuracy, and high cost of character\nsegmentation annotations. Recently, segmentation-free methods based on\nconnectionist temporal classification (CTC) and attention mechanism, have\ndominated the field of HCTR. However, people actually read text character by\ncharacter, especially for ideograms such as Chinese. This raises the question:\nare segmentation-free strategies really the best solution to HCTR? To explore\nthis issue, we propose a new segmentation-based method for recognizing\nhandwritten Chinese text that is implemented using a simple yet efficient fully\nconvolutional network. A novel weakly supervised learning method is proposed to\nenable the network to be trained using only transcript annotations; thus, the\nexpensive character segmentation annotations required by previous\nsegmentation-based methods can be avoided. Owing to the lack of context\nmodeling in fully convolutional networks, we propose a contextual\nregularization method to integrate contextual information into the network\nduring the training stage, which can further improve the recognition\nperformance. Extensive experiments conducted on four widely used benchmarks,\nnamely CASIA-HWDB, CASIA-OLHWDB, ICDAR2013, and SCUT-HCCDoc, show that our\nmethod significantly surpasses existing methods on both online and offline\nHCTR, and exhibits a considerably higher inference speed than\nCTC/attention-based approaches.\n