FPRNet: End-to-End Full-Page Recognition Model for Handwritten Chinese Essay

Handwritten Chinese Essay Recognition (HCER) is a special branch of handwritten Chinese text recognition with great interest. In a naive way, it can be firstly segmented into text lines or even characters, followed by a text line or character recognition step. Instead, we propose an end-to-end recognition model named FPRNet which directly runs on full-page images in light of the segmentation-free strategy. Our well-designed model can extract text from a full-page image only supervised with text labels and adapt better to authentic noisy images. Besides, we propose an effective dimensionality reduction mechanism based on reshape operation to bridge features between 2D and 1D without information loss. Moreover, we propose an order-align strategy to mitigate the decoding confusion caused by skewness. Experiments are conducted on real-world essay images. Our model achieves a 5.83% character error rate (CER), which is comparable with the state-of-the-art approaches.