Paper Overview
- Field: Computer Vision
- Authors: Md. Sultan Al Rayhan, Maheen Islam
- Published: 2025-05-09
- arXiv: 2505.07237
- Combines class-conditional diffusion modeling with classifier guidance to synthesize high-quality handwritten compound character samples.
- Introduces Squeeze-and-Excitation enhanced residual blocks in the diffusion model's U-Net backbone to improve generation quality.
- Applies a confidence-based filtering mechanism where a pretrained classifier acts as a quality gate, retaining only highly class-consistent synthetic samples.
- Merges filtered synthetic images with original training data to retrain multiple classification architectures.
Summary
Recognition of handwritten Bangla compound characters remains a challenging problem due to complex character structures, large intra-class variation, and limited availability of high-quality annotated data. Existing Bangla handwritten character recognition systems often struggle to generalize across diverse writing styles, particularly for compound characters containing intricate ligatures and diacritical variations.
Method
The authors propose a confidence-guided diffusion augmentation framework for low-resolution Bangla compound character recognition:
Results
Experiments on the AIBangla compound character dataset show consistent performance improvements across ResNet50, DenseNet121, VGG16, and Vision Transformer architectures. The best-performing model achieves 89.2% classification accuracy, substantially surpassing previously published AIBangla benchmarks.
Conclusion
Quality-aware diffusion augmentation can effectively improve handwritten character recognition performance in low-resource script domains.
--- *Auto-collected on 2026-05-13*