Abstract
Scene text recognition plays an important role in many computer vision applications. The small size of available public available scene text datasets is the main challenge when training a text recognition CNN model. In this paper, we propose a CNN based Chinese text recognition algorithm. To enlarge the dataset for training the CNN model, we design a synthetic data engine for Chinese scene character generation, which generates representative character images according to the fonts use frequency of Chinese texts. As the Chinese text is more complex, the English text recognition CNN architecture is modified for Chinese text. To ensure the small size nature character dataset and the large size artificial character dataset are comparable in training, the CNN model are trained progressively. The proposed Chinese text recognition algorithm is evaluated with two Chinese text datasets. The algorithm achieves better recognize accuracy compared to the baseline methods.
Introduction
While English scene text recognition has advanced significantly, Chinese scene text recognition remains a formidable challenge due to two primary factors: structural complexity and data scarcity. Unlike the 26-letter English alphabet, standard Chinese character sets contain over 10,000 entries with intricate stroke patterns. Furthermore, deep learning models require millions of labeled examples, yet public Chinese scene text datasets are often limited to a few thousand images. This paper addresses these challenges through a specialized synthetic data engine and a tailored CNN architecture.
Methodology
Synthetic Data Engine
To overcome the lack of labeled data, we developed a synthetic data engine that generates high-quality training images mimicking real-world conditions. The generation process involves font selection weighted by real-world frequency, stylization (colors, borders, shadows), and the addition of environmental noise (background textures and Gaussian blur).
CNN Architecture and Progressive Training
Standard CNNs for English text often use small convolutional filters that fail to capture the complex stroke relationships in Chinese characters. We modified the architecture by enlarging these windows to improve feature extraction.
To effectively utilize both synthetic and real data, we employed a Progressive Training strategy:
- Initial Phase: Training on a large synthetic dataset (95% artificial) to learn basic character shapes.
- Fine-tuning Phase: Training on a balanced mix of synthetic and natural images to adapt the model to real-world visual noise without overwhelming it with artificial data.
Experimental Results
The proposed model (CNN-7) was evaluated against baseline methods on Ren’s Dataset (natural) and Zhou’s Dataset (mixed).
| Model | Ren’s Dataset (Natural) | Zhou’s Dataset (Mixed) |
|---|---|---|
| CNN-7 (A+S) | 0.73 | 0.76 |
| CNN-9 (A+S) | 0.73 | 0.75 |
| ABBYY (Standard OCR) | 0.40 | 0.41 |
The results demonstrate that combining artificial (A) and scene (S) data nearly doubles recognition accuracy compared to using natural data alone (0.38 on Ren's dataset).
Discussion and Conclusion
Our findings indicate that for Chinese text, deeper networks are not necessarily superior; a 7-layer model with larger filters outperformed a 9-layer model with smaller filters. This suggests that capturing stroke relationships is more critical than depth for complex scripts. While the current system focuses on the 1,200 most common characters, future work will aim to expand the vocabulary and optimize computational efficiency for real-time applications.
Summary updated