Reproduce “Entropy-Aware Structural Alignment for Zero-Shot Handwritten Chinese Character Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qiuming Luo, Tao Zeng, Feng Li, Heming Liu, Rui Mao, Chang Kong
The Entropy-Aware Structural Alignment Network (EASA) improves zero-shot and few-shot Handwritten Chinese Character Recognition by integrating information entropy to weight radicals and employing multi-granularity structural modeling with cross-modal attention. It achieved 55.04% Top-1 accuracy on unseen characters in ICDAR 2013 and demonstrated high data efficiency in few-shot scenarios.
Key insights
- Incorporating information entropy allows the model to dynamically prioritize the most discriminative radicals, enhancing character representation and reducing ambiguity caused by ubiquitous, low-information radicals.
- Explicitly modeling the hierarchical 2D structure of Chinese characters at both global (parent-centric) and local (child-centric) levels is crucial for robustly aligning noisy handwritten visual features with structured semantic prototypes.
- Deep cross-modal attention, combined with a Top-K semantic feature fusion strategy, provides a more resilient and accurate alignment mechanism against handwriting distortions compared to shallow similarity metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.