Chinese character-level writer identification using path signature feature, DropStroke and deep CNN

Weixin Yang, Lianwen Jin, Manfei Liu

Introduction

Writer identification is a critical task in digital forensics and security. Traditional systems often require large amounts of text—paragraphs or full pages—to reliably identify a writer. However, in many real-world scenarios, only a few characters may be available. Identifying a writer from a single Chinese character is particularly challenging due to the vast character set (over 50,000 classes) and the need for the system to be "text-independent," focusing on the writer's style rather than the character's shape.

This paper presents an end-to-end system that achieves high accuracy by treating handwriting as a temporal path and employing a novel data augmentation technique called DropStroke.

Methodology

The proposed system integrates three core components: Path Signature features, a Deep Convolutional Neural Network (DCNN), and the DropStroke augmentation method.

Path Signature Features

Handwriting is treated as a path in time rather than a static image. The researchers use "iterated integrals" or path signatures to represent the trajectory. This method captures subtle geometric details like displacement and curvature, which are often lost in standard bitmap representations.

Deep CNN Architecture

The system utilizes a DeepCNet architecture, consisting of six convolutional layers and two fully-connected layers. This design is specifically optimized to handle the spatial sparsity inherent in handwritten characters.

DropStroke Augmentation

To address data scarcity and force the model to learn "micro-styles," the authors introduced DropStroke. Similar to "Dropout" in neural networks, DropStroke randomly removes individual strokes from a character during training. This process generates thousands of synthetic training samples from a single character and prevents the network from over-relying on global character structure.

Overview of the proposed writer identification system, illustrating the integration of DropStroke augmentation, Path Signature feature extraction, and the Deep CNN architecture.

Dataset and Experimental Setup

The study utilized the CASIA-OLHWDB1.0 dataset, featuring 3,866 character classes from 420 writers. To ensure a text-independent evaluation, only 200 samples per writer were used for training, while the remaining 3,666 samples were reserved for testing.

Results and Analysis

The experimental results demonstrate that temporal information and DropStroke are vital for high-performance writer identification.

Performance Metrics

Using the optimal configuration (Sign.3 + DropStroke), the system achieved a 55.45% Top-1 accuracy for identifying one of 420 writers from just a single character. Accuracy improves significantly as more characters are provided:

  • 3 characters: 80.33%
  • 10 characters: 97.58%
  • 20 characters (approx. one line): 99.52%

Impact of Path Signatures and DropStroke

The researchers found that higher-level path signatures (Sign.3) consistently outperformed simple offline bitmaps (Sign.0). Furthermore, the addition of DropStroke significantly lowered the test error rate across all signature levels, as shown in the comparative performance plots.

Test error rates across training iterations, comparing different Path Signature levels (a) and the performance boost provided by DropStroke (b).

Conclusion and Future Work

The combination of Path Signature features and DropStroke allows for highly accurate, text-independent writer identification from minimal data. While the current system requires "online" data (temporal coordinate sequences), it sets a new benchmark for character-level identification. Future work may involve adapting these techniques for "offline" scanned images and exploring the utility of DropStroke in general character recognition or gender identification tasks.

Summary updated

BibTeX
@inproceedings{yang2015chinese,
  author = {Weixin Yang and Lianwen Jin and Manfei Liu},
  title = {{Chinese character-level writer identification using path signature feature, DropStroke and deep CNN}},
  booktitle = {ICDAR},
  year = {2015},
  url = {https://openreview.net/forum?id=zPR1drmeoa},
}
Figure