TL;DR
This paper introduces a novel deep Siamese neural network approach for handwritten Chinese character recognition (HCCR) that treats recognition as a template matching problem. This method excels at recognizing characters not seen during training, addressing the "open set" problem in HCCR.
What Problem Does This Paper Address?
Handwritten Chinese Character Recognition (HCCR) is challenging due to the vast number of characters, variations in writing styles, and the difficulty of obtaining sufficient training data for every character. Traditional methods often struggle with "closed set" recognition (where all test characters are present in the training set) and cannot generalize to new, unseen characters. This paper tackles the "open set" HCCR problem, aiming to recognize characters that were not part of the initial training data, a crucial capability for real-world applications with ever-expanding character sets.
Methods
The core of this work is a deep Siamese neural network adapted for template matching. Instead of a traditional multi-class classifier, the network learns to predict the similarity between an input handwritten character and a set of machine-printed character templates.
Here's a breakdown:
- Siamese Network Architecture: The network consists of two identical subnetworks that share weights. Each subnetwork processes an input (either a handwritten character or a template) and outputs a feature vector.
- Template Matching: The similarity between the input handwritten character and a template is measured using the L₁ distance between their respective feature vectors.
- Binary Classification Training: The Siamese network is trained as a binary classifier. Pairs of (handwritten character, template) are fed into the network, and the output is trained to predict 1 if they match and 0 if they don't. This training process optimizes the network to learn discriminative features.
- Open-Set Recognition: Once trained, the network can recognize unseen characters. For a given handwritten character, its features are extracted and compared against features of all potential templates. The template with the highest similarity score determines the recognized character. This process doesn't require retraining for new characters.
Datasets
- Training: CASIA-HWDB1.0 and CASIA-HWDB1.1 datasets were used for training the Siamese network. These datasets contain approximately 2.67 million samples from 720 writers.
- Evaluation: The ICDAR-2013 offline HCCR competition datasets were used for evaluation. These datasets contain approximately 0.22 million samples from 60 writers.
Key Insights and Ideas
- Template Matching as a Solution for Open Sets: Framing HCCR as template matching, rather than a direct multi-class classification, is a key innovation. This allows the model to generalize to characters not seen during training without explicit retraining.
- Data Augmentation through Pair Generation: Training with sample pairs (positive and negative) naturally augments the dataset. For a c-class task with n samples per class, nc² sample pairs can be generated, significantly increasing the effective training data size.
- Fixed Model Size: Unlike traditional classifiers where model size scales with the number of classes, this template matching approach has a fixed model size, regardless of the number of characters to be recognized.
- Discriminative Feature Learning: The Siamese network's training objective forces it to learn highly discriminative features that capture subtle differences between characters, crucial for accurate similarity prediction.
Key Results
- Promising Generalization: Experiments on the ICDAR-2013 datasets demonstrated that the proposed method has a "very promising generalization ability to the new classes that never appear in the training set."
- Open-Set Performance: While not explicitly quantified with a single "open-set accuracy" metric in the provided text, the paper highlights that the method can recognize characters not seen during training.
- Comparison to State-of-the-Art (Closed Set): In a closed-set scenario (all characters seen during training), the template-matching classifier achieved an accuracy of 92.31% on the ICDAR-2013 datasets, which is lower than some state-of-the-art methods (e.g., DirectMap+CNN at 96.95%). The authors attribute this to the inherent difficulty of learning similarity metrics for similar handwritten characters.
- Error Analysis: Figure 5 provides examples of top-10 failure cases, showing visually similar characters being confused (e.g., 丁 vs. 了, 且 vs. 旦).
Limitations and Future Work
- Lower Accuracy in Closed Set: The authors acknowledge that their template-matching classifier achieves lower accuracy in a closed-set scenario compared to other state-of-the-art methods. This is attributed to the difficulty of learning fine-grained similarity metrics for visually similar handwritten characters.
- Future Work: The authors plan to:
- Focus on better understanding the error cases.
- Further improve the model.
- Reduce the performance gap with state-of-the-art methods.
- Explore the application to "open set" handwritten Chinese character recognition, which they claim is a novel contribution of their work.
Summary updated