On the Behavior of Contrastive Regularization in Improving Chinese Text Recognizer

Dekang Liu, Tianlei Wang, Huanqiang Zeng, Jiuwen Cao

Abstract: The dense representation space in Chinese scene text recognition (STR) makes discriminating between categories highly challenging, because of the large candidate category set. Mainstream STR methods have achieved remarkable advancements by leveraging linguistic knowledge to implicitly address this challenge. In this paper, inspired by the correlation between recognizer performance and the distributional properties of character representations, as well as the inherent consistency between this correlation and supervised contrastive learning (SupCon), we thoroughly investigate how to integrate SupCon with an STR model to alleviate this challenge, and elucidate some dynamic behaviors underlying the performance improvements. Specifically, we analyze the SupCon-STR models instantiated with different projectors and evaluate their distributional properties through metrics, including intra-class compactness, inter-class separability, and feature redundancy, while assessing performances that involve in-domain accuracy and cross-domain recognition generalization. The main results reveal how the temperature $\tau$ and projectors affect the representation distribution, and highlight that suitable intra-class compactness and sufficient inter-class separability are key factors for delivering competitive performances in both in-domain and cross-domain STR scenarios. Moreover, these results also provide valuable insights into the design of SupCon-STR architectures for diverse resource constraints. Taking existing Chinese STR models as baselines, and combining SupCon-STR with them, the average improvements in cross-domain recognition performance are over 5% across 7 testing datasets. A new state-of-the-art accuracy of 77.19% on the Chinese Scene benchmark is also established.

External IDs:doi:10.1109/tmm.2025.3604892