Abstract
Offline Handwritten Chinese Character Recognition (HCCR) faces significant challenges due to the vast vocabulary and structural complexity of Chinese characters. Traditional methods treat characters as monolithic classes, leading to data sparsity and an inability to recognize unseen characters. This paper introduces DenseRAN (Dense Radical Analysis Network), which treats HCCR as an image captioning task. By decomposing characters into their constituent radicals and two-dimensional structures, DenseRAN achieves state-of-the-art accuracy on standard benchmarks and demonstrates a unique zero-shot capability for recognizing characters not present in the training set.
Introduction
The primary hurdle in HCCR is the sheer scale of the Chinese language, which contains tens of thousands of characters. Conventional deep learning models typically employ a large softmax layer where each character is a unique class. This approach suffers from two major drawbacks: first, the difficulty of collecting sufficient handwritten samples for every character; and second, the failure to recognize "unseen" characters during inference.
To address these issues, the authors leverage the inherent compositionality of Chinese characters. Most characters are composed of a finite set of approximately 500 radicals arranged in specific 2D spatial configurations (e.g., left-right, top-bottom). By learning to identify these components and their structural relationships, a model can theoretically "read" any character, even those it has never encountered as a whole.
Methods: The DenseRAN Architecture
DenseRAN utilizes an encoder-decoder framework to translate character images into structural "captions."
Encoder and Decoder
The encoder is a Densely Connected Convolutional Network (DenseNet), chosen for its efficiency in feature reuse and robust gradient flow. It extracts high-level visual features from the input handwritten image. The decoder is a Gated Recurrent Unit (GRU) that generates a sequence of radicals and structural markers.
Structural Decomposition and Attention
Characters are represented as strings that describe their internal geometry. For example, a character with a left-right structure is decoded as: a { radical-1, radical-2 }, where 'a' denotes the structure. A coverage-based spatial attention mechanism allows the decoder to focus on specific image regions corresponding to each radical as it generates the sequence.
Experimental Results
The model was evaluated using the CASIA-HWDB (1.0, 1.1, and 1.2) and ICDAR-2013 datasets.
Performance on Seen Characters
On the ICDAR-2013 benchmark, DenseRAN achieved an accuracy of 96.66%, surpassing the human performance benchmark of 96.13% and a standard whole-character DenseNet (95.90%). This represents an 18.54% relative reduction in the Character Error Rate (CER).
Zero-Shot Recognition
The most compelling result is the model's performance on 3,277 character classes from CASIA-HWDB 1.2 that were excluded from the training set. While traditional classifiers score 0% on such tasks, DenseRAN achieved 40.82% accuracy. This demonstrates that the model has learned the underlying "grammar" of Chinese characters, allowing it to parse novel combinations of known radicals.
Discussion and Limitations
Despite its success, DenseRAN faces challenges with structural ambiguity in highly cursive or "scribbled" handwriting, where boundaries between radicals blur. Additionally, characters with very subtle visual differences (e.g., '已' vs '己') can occasionally confuse the attention mechanism. Future work will focus on enhancing the attention model's robustness and exploring methods to better handle diverse handwriting styles.
Conclusion
DenseRAN represents a paradigm shift in HCCR by moving from monolithic classification to compositional analysis. By mimicking the human process of parsing characters into radicals and structures, the model achieves superior accuracy and provides a path toward recognizing the full breadth of the Chinese written language.
Summary updated