Diverse Feature Generation for Zero-Shot Chinese Character Recognition Reference

Song-Liang Pan, Kunchi Li, Da-Han Wang, Xu-Yao Zhang, Jiantao Liu, Shunzhi Zhu

Abstract: Zero-shot Chinese Character Recognition (ZSCCR) aims to identify unseen Chinese characters not included in the training set. Traditional ZSCCR approaches often rely on predicting radicals to bridge the gap between seen and unseen classes. However, in zero-shot learning (ZSL), these models are restricted to learning semantics from seen class data, resulting in a severe bias problem. Recently, synthesizing unseen class features using semantic information in ZSL to alleviate data imbalance has become a trend. However, few studies have explored applying this strategy to ZSCCR, and maintaining diversity in the generated features remains a significant challenge. To address these challenges, this work proposes a Diverse Feature Generation (DFG) framework for ZSCCR. Specifically, DFG synthesizes features using Ideographic Description Sequences (IDS) for unseen characters, solving the lack of training data for unseen classes and mitigating recognition bias in ZSCCR. DFG introduces a Hybrid Semantic Embedding (HSE) strategy to enrich input semantics and employ multiple generative sub-networks to produce diverse features with varying semantics. Additionally, to further enhance feature diversity, a Diversity-Increase Loss (DI-loss) is proposed, which encourages sub-networks to recognize features with distinct semantics and increases the entropy of generated features within the same class. A Prediction-Level Feature Collaboration Loss (PLFC-loss) is also introduced to encourage collaboration among sub-networks during training, helping to mitigate domain shift and further enhance performance. Experimental results demonstrate that the proposed method performs well in generalized ZSCCR settings across benchmark datasets.