Recognizing Chinese Texts with Multi-width Feature Extractor and Attention-based Fusion
Pu Cao · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
Recognizing texts plays an important role in optical characters recognition (OCR). Although the previous text recognition methods have made great progress, most of them focus on Latin characters, while few are about Chinese characters. There are two main differences between Chinese characters and Latin characters. First, the amount of Chinese characters is much larger than that of Latin characters. Second, the width of Chinese characters varies from character, while that of Latin characters is stable. This paper proposes a multi-width text recognition method to solve the two challenges in Chinese text recognition. A multiple feature extractor module is introduced to obtain multiple features of characters in a Chinese task. Besides, an attention-based fusion module is employed to dynamically fuse the multiple features. We conduct experiments on a popular Chinese dataset, and the results demonstrate that our model is superior to other state-of-the-art (SOTA) baselines.