Improved deep convolutional neural network for online handwritten Chinese character recognition using domain-specific knowledge

Weixin Yang, Lianwen Jin, Zecheng Xie, Ziyong Feng

Introduction

Online Handwritten Chinese Character Recognition (HCCR) has long been a challenging frontier in pattern recognition. Unlike offline recognition, which deals with static images, online recognition captures the temporal sequence of pen strokes. While Deep Convolutional Neural Networks (DCNNs) have significantly advanced the field, they often treat handwriting as simple 2D bitmaps, potentially ignoring decades of linguistic and structural research. This paper presents a novel framework that fuses traditional domain-specific expertise with modern deep learning to achieve state-of-the-art accuracy and computational efficiency.

Methodology

The authors propose a streamlined DCNN architecture that is significantly more efficient than its predecessors, utilizing 3.8 million parameters compared to the 5.9 million found in leading models like DeepCNet. The core innovation lies in the representation of input data and the strategy for model ensemble.

1. Domain-Specific Feature Integration

Rather than relying on raw pixel data, the network is fed multiple specialized data channels that "pre-digest" the structural rules of Chinese calligraphy:

  • Path Signatures: Mathematical iterated integrals that capture the underlying geometry of the pen's movement.
  • 8-Directional Features: Explicit maps highlighting stroke orientations, providing the network with a clear sense of character structure.
  • Imaginary Strokes: These represent the "pen-up" trajectories—the movement of the hand between actual ink strokes. This captures the rhythmic flow and intent of the writer, which is often lost in static representations.
  • Non-linear Normalization (NLN): A technique used to adjust for variations in stroke spacing and writing styles.

2. Hybrid Serial-Parallel (HSP) Ensemble

To balance high accuracy with real-time performance, the authors introduced the Hybrid Serial-Parallel (HSP) ensemble strategy. Traditional ensembles run multiple models simultaneously and average their results, which is computationally expensive. The HSP strategy functions as a triage system:

  1. An input is first processed by a single, efficient DCNN.
  2. If the model's confidence exceeds a threshold (probability > 0.99), the system outputs the result immediately.
  3. If the confidence is low, the sample is passed to specialized models.
  4. Only the most difficult cases are processed by the full ensemble for a final vote.

Experimental Evaluation

The proposed system was evaluated using two benchmark datasets from the Institute of Automation of the Chinese Academy of Sciences: CASIA-OLHWDB1.0 (3,740 classes) and CASIA-OLHWDB1.1 (3,755 classes).

Performance Metrics

  • Accuracy: The model achieved 97.20% on DB1.0 and 96.87% on DB1.1, outperforming previous benchmarks such as DeepCNet (96.42%) and DLQDF (94.85%).
  • Error Reduction: By incorporating domain-specific knowledge, the authors reduced the relative error rate by 42% compared to a baseline DCNN using only raw bitmaps.
  • Efficiency: The HSP ensemble strategy proved to be 4.5 times faster than standard averaging ensembles, averaging 6.22ms per sample. This speedup is primarily because approximately 84% of samples are correctly identified by the first model in the sequence.

Key Insights

The Value of Pre-Digested Features

A significant finding was that providing explicit directional features improved accuracy even when using Path Signatures (which theoretically already contain that information). This suggests that DCNNs perform better when critical features are explicitly highlighted rather than left for the network to discover from scratch.

The Importance of "Imaginary" Data

The inclusion of imaginary strokes—the path the pen takes while lifted—provided a substantial boost to recognition performance. This indicates that the temporal "rhythm" and the transition between strokes are as discriminative as the strokes themselves.

Architectural Observations

Interestingly, Non-linear Normalization (NLN) provided less benefit than expected. The authors hypothesize that the max-pooling layers within the DCNN are inherently capable of handling the spatial variations that NLN was traditionally designed to correct, rendering the old-school normalization somewhat redundant in a deep learning context.

Conclusion

By combining the structural insights of traditional handwriting research with the feature-learning power of DCNNs, this research sets a new standard for Online HCCR. The Hybrid Serial-Parallel ensemble strategy further demonstrates that state-of-the-art accuracy does not have to come at the cost of computational overhead, making high-performance handwriting recognition viable for mobile and real-time applications.

Summary updated

BibTeX
@inproceedings{yang2015improved,
  author = {Weixin Yang and Lianwen Jin and Zecheng Xie and Ziyong Feng},
  title = {{Improved deep convolutional neural network for online handwritten Chinese character recognition using domain-specific knowledge}},
  booktitle = {ICDAR},
  year = {2015},
  url = {https://openreview.net/forum?id=W1v7ZKHbZSji},
}
Figure