Chinese/English mixed Character Segmentation as Semantic Segmentation.

OCR character segmentation for multilingual printed documents is difficult due to the diversity of different linguistic characters. Previous approaches mainly focus on monolingual text and are not suitable for multi-lingual cases. In this work, we particularly tackle the Chinese/English mixed case by reframing it as a semantic segmentation problem. We take advantage of the successful architecture called fully convolutional networks (FCN) in the field of semantic segmentation. As a deep architecture, FCN can automatically learn useful features without traditional feature engineering. Given wide enough receptive field, it can utilize the necessary context around a position to better determinate whether this is a splitting point or not. Trained on synthesized samples with simulated random disturbances, FCN can effectively split characters without any hand-crafted features. The experimental results show that our model significantly outperforms the previous methods. It is able to generalize from simulated disturbances to real-world disturbances, generalize from one text content style to another, generalize from seen font styles to unseen ones, and correctly handle disconnected structures and touching characters.