Recognition of Online Handwritten Chinese Texts in Any Writing Direction via Stroke Classification Based Over-Segmentation

Yi Chen, Heng Zhang, Min-Si Ren, Cheng-Lin Liu

Abstract: Online handwritten text recognition technology has been increasingly applied in intelligent touch-based and pen-based devices. Current mainstream methods are mostly designed for horizontally written texts, thus are difficult to handle texts in any writing direction. This paper proposes a recognition framework based on over-segmentation which is applicable to text recognition of any writing direction. It divides text line inclination styles into two cases: texts with the entire line rotated and texts with the line direction rotated while keeping the characters upright. A text line inclination style classification module is introduced in the preprocessing stage to classify these two cases. The former case can be recognized using a horizontal text line recognizer after rotation correction. For the latter case, an improved over-segmentation algorithm is designed based on stroke classification using bidirectional long short-term memory networks (BiLSTM) to achieve text recognition in any writing direction. Experimental results demonstrate that the proposed method is capable of text recognition in any writing direction and achieves highly competitive results on the CASIA-OLHWDB and ICDAR2013-Online datasets.