Hierarchical Discriminative Feature Aggregation Based Offline Handwritten Chinese Text Recognition

Connectionist temporal classification (CTC) based methods for offline handwritten Chinese text recognition (HCTR) have achieved remarkable progress recently due to their superiority in the robustness to varying sequence lengths and inference efficiency. However, CTC-based models still grapple with the attenuation of the feature representation degradation because of the collapse of height information of text images. To alleviate this issue, in this paper, we propose utilizing deep supervision technology and hierarchical feature aggregation to improve the CTC-based recognizers and construct a novel HCTR model, namely DSHctr. Concretely, DSHctr deploys multiple classifiers to directly supervise and optimize discriminative multi-scale features. Then these hierarchical discriminative are aggregated for the final recognition. The proposed DSHctr model exhibits remarkable performance on the publicly available ICDAR-2013 handwriting competition dataset, achieving an accuracy rate of 94.29% and a correct rate of 94.49% without a language model, surpassing its state-of-the-art counterparts. Extensive qualitative and quantitative experiments highlight the effectiveness of the proposed hierarchical discriminative feature aggregation.