Structural Recognition of Handwritten Chinese Characters Using a Modified Part Capsule Auto-encoder

Xinjian Wu, Xiang Ao, Rui-Song Zhang, Cheng-Lin Liu

Abstract: Handwritten Chinese character recognition has achieved high accuracy using deep neural networks (DNNs), but the structural recognition (which offers structural interpretation, e.g., stroke and radical composition) is still a challenge. Existing DNNs treat character image as a whole and perform classification end-to-end without perception of the structure. They need a large amount of training samples to guarantee high generalization accuracy. In this paper, we propose a method for structural recognition of handwritten Chinese characters based on a modified part capsule auto-encoder (PCAE), which explicitly considers the hierarchical part-whole relationship of characters, and leverages extracted structural information for character recognition. Our PCAE is improved based on stacked capsule auto-encoder (SCAE) so as to better extract strokes and perform classification. By the modified PCAE, the character image is firstly decomposed into primitives (stroke segments), with their shape and pose information decoupled. The transformed primitives are aggregated into higher-level parts (strokes) guided by prior knowledge extracted from writing rules. This process enhances interpretability and improves the discrimination ability of features. Experimental results on a large dataset demonstrate the effectiveness of our method in both Chinese character recognition and stroke extraction tasks.