Handwriting Trajectory Reconstruction Using Spatial-Temporal Encoder-Decoder Network

Chinese handwriting characters have complex strokes and various writing styles, which makes it difficult to reconstruct handwriting. Aiming at this problem, we propose a handwriting reconstruction method based on a spatial-temporal encoder-decoder network with constrains. Different from other models that generate trajectory coordinates through a fully connected network, the method proposed in this paper outputs heat map sequence. The model is consists of three modules: key point detection module, spatial encoder-decoder module and reconstruction constraint module. The key point detector module and the spatial encoder part of encoder-decoder module are composed of a full convolutional network. The former generates heat maps of all key points which is a branch of the spatial encoder, and the mainly encoding the spatial information of each position on the offline image. The temporal decoder module is composed of a GRU network and an MLP network. Finally, we combine temporal information and reconstruction constraints to generate the final sequence. At each time, the features encoding by the spatial encoder module are combined with the features at the previous time that generate a corresponding heat map. The main contribution of the work of this paper is to propose a method that more suitable for handwriting reconstruction of Chinese handwritten characters. Experimental results show that the CT [6] accuracy of our method has already reached 87.6% on OLHWDB1.1 dataset.