DBNet for Robust Text Detection in Challenging Medical Packaging Images
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Text detection is a fundamental task in computer vision, supporting a wide range of applications such as scene understanding, document analysis, and intelligent information management. In the context of pharmaceutical packaging, detecting functional indication text is particularly challenging due to its small size, low contrast, complex background patterns, and diverse font styles. Traditional OCR-based approaches often struggle in these scenarios, leading to incomplete or inaccurate extraction. To address these issues, we propose a robust text detection framework based on DBNet (Differentiable Binarization Network). The core mechanism of differentiable binarization enables adaptive separation of text from cluttered backgrounds, which significantly improves detection robustness. Furthermore, we adopt a lightweight ResNet18 backbone to enhance efficiency while maintaining high accuracy, making the model well-suited for real-world deployment. A custom dataset of real-world medicine box images was constructed, covering diverse lighting conditions, distortions, and viewing angles. Comprehensive experiments demonstrate that our DBNet with ResNet18 achieves the best trade-off between accuracy and speed, with Precision = 0.9289, Recall = 0.8213, and F1-score = 0.8718, outperforming baseline models. These results confirm that the proposed method is not only accurate but also computationally efficient. This work highlights the potential of DBNet-based approaches for reliable text detection in complex scenarios, providing a solid foundation for automated pharmaceutical information management systems.
Fields of Study & Subjects
Handwritten Text Recognition Techniques
Advanced Neural Network Applications
Vehicle License Plate Recognition
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 897–901
- Publication date
- 28 January 2026
- Language
- EN