🤖 AI

Confidence-Guided Diffusion Augmentation for Enhanced Bangla Compound Character Recognition

This paper proposes a confidence-guided diffusion augmentation framework that synthesizes high-quality handwritten Bangla compound characters using class-conditional diffusion models with Squeeze-and-Excitation enhancements and a filtering mechanism, achieving a new state-of-the-art accuracy of 89.2% on the AIBangla dataset.

Original authors: Md. Sultan Al Rayhan, Maheen Islam

Published 2026-05-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Md. Sultan Al Rayhan, Maheen Islam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to read handwritten notes in Bangla. This is a tricky task because Bangla letters often combine to form complex, "compound" shapes that look very different depending on who wrote them. It's like trying to recognize a friend's face when they are wearing different hats, scarves, and sunglasses every day.

The main problem the researchers faced was a lack of "practice material." To teach a computer well, you need thousands of examples, but for these specific Bangla shapes, there just weren't enough high-quality, labeled images available.

Here is how the authors solved this problem, broken down into simple steps:

1. The "Art Student" (The Diffusion Model)

Instead of just copying existing images, the team taught a computer program called a Diffusion Model to act like a creative art student.

How it works: Imagine a student who starts with a blank canvas covered in static noise (like TV snow). They slowly remove the noise, step by step, until a clear picture of a handwritten letter emerges.
The Twist: The researchers didn't just let the student draw whatever they wanted. They gave the student a specific "assignment" (a specific letter) and a strict "teacher" (classifier guidance) to ensure the drawing looked exactly like that letter.
The Upgrade: To make the drawings even better, they added a special tool called Squeeze-and-Excitation blocks. Think of this as giving the art student a magnifying glass and a highlighter, helping them focus on the most important details of the letter so the final drawing is crisp and accurate.

2. The "Strict Gatekeeper" (Confidence Filtering)

The art student is creative, but sometimes they might draw a messy or confusing letter that looks nothing like the real thing. If you feed these bad drawings to the main computer, it will get confused and learn the wrong things.

To fix this, the team introduced a Gatekeeper:

Before the new drawings are added to the training pile, a pre-trained computer (the Gatekeeper) looks at them.
If the Gatekeeper is 90% sure that the drawing is the correct letter, it lets it through.
If the Gatekeeper is unsure or thinks the drawing is garbage, it throws it away.
This ensures that only the "gold standard" fake images get mixed in with the real ones.

3. The "Study Group" (Retraining)

Once they had a pile of real handwritten letters and a filtered pile of high-quality fake letters, they mixed them together. They then used this massive, improved study group to retrain four different types of computer "brains" (called ResNet50, DenseNet121, VGG16, and Vision Transformer).

The Result

The experiment was a huge success.

The Goal: Recognize complex Bangla compound characters.
The Outcome: The computer models got significantly smarter. The best model (VGG16) achieved 89.2% accuracy.
The Comparison: This is a big jump from previous records, beating the best existing benchmarks by a wide margin.

Why This Matters (According to the Paper)

The paper emphasizes that this method works well even with low-resolution images (small, 32x32 pixel pictures). This means the system is efficient and could potentially run on devices that don't have super-powerful computers, making it practical for real-world document scanning.

In a nutshell: The researchers taught a computer to draw fake Bangla letters, used a strict filter to keep only the perfect drawings, and then used those perfect fakes to train other computers to read handwriting much better than before.

Technical Summary: Confidence-Guided Diffusion Augmentation for Enhanced Bangla Compound Character Recognition

Problem Statement
Handwritten Bangla compound character recognition faces significant hurdles due to the script's complex structural ligatures, high intra-class variation across diverse writing styles, and a scarcity of high-quality annotated data. Existing systems often struggle to generalize, particularly when dealing with intricate compound characters, and many rely on computationally expensive high-resolution inputs that limit deployment in resource-constrained environments. While generative models like GANs have been explored, they often suffer from instability and mode collapse. Furthermore, the application of diffusion models to low-resource Indic scripts, specifically for compound character synthesis, remains underexplored.

Methodology
The authors propose a confidence-guided diffusion augmentation framework designed to synthesize high-quality handwritten samples to augment limited training data. The workflow consists of four primary stages:

Diffusion Model Architecture: The core generative model utilizes a U-Net backbone enhanced with Squeeze-and-Excitation (SE) residual blocks to improve feature representation. The architecture incorporates time embeddings, learned class embeddings, and multi-head attention mechanisms. The model is trained to predict injected noise at each diffusion timestep.
Classifier Guidance: To ensure the generated images align with specific target classes, the framework employs a class-conditional diffusion model guided by an auxiliary classifier. This classifier, trained on noisy images, provides gradients during the sampling process to steer the generation toward the desired class.
Confidence-Based Filtering: A critical component of the framework is a quality control mechanism. Pre-trained classifiers act as "quality gates" to filter the generated synthetic images. Only samples with a confidence score of $\ge$ 0.90 for their intended class are retained. This step is designed to prevent low-quality or ambiguous synthetic samples from degrading the performance of downstream classifiers.
Data Fusion and Retraining: The filtered synthetic images are fused with the original real-world training data. This augmented dataset is then used to retrain multiple classification architectures, including ResNet50, DenseNet121, VGG16, and Vision Transformer (ViT).

The entire pipeline operates on low-resolution 32×32 grayscale images, aiming for computational efficiency while maintaining suitability for real-world document analysis.

Key Contributions
The paper outlines five specific contributions:

The proposal of a confidence-guided diffusion augmentation framework specifically for Bangla compound character recognition.
The integration of SE-enhanced residual blocks within the diffusion U-Net backbone to enhance synthetic image quality.
The development of a classifier-confidence-based filtering strategy to rigorously remove low-quality synthetic samples prior to retraining.
The demonstration of consistent performance improvements across diverse classifier architectures (CNNs and Transformers) on the AIBangla benchmark.
The establishment of a new benchmark performance record for the AIBangla compound character recognition task.

Results
Experiments were conducted on the AIBangla dataset, which contains 249,911 handwritten compound character images across 171 classes.

Synthesis Quality: The diffusion model achieved a Fréchet Inception Distance (FID) of 24.63. The application of confidence-based filtering further improved the FID scores across all filtering architectures (e.g., ViT filtering reduced FID to 24.04), indicating that the retained synthetic samples were of higher distributional quality.
Classification Performance: Retraining on the augmented dataset yielded consistent improvements across all tested models.
- VGG16 emerged as the best-performing model, achieving 89.2% accuracy (up from a baseline of 88.2%), with a precision of 0.8470 and recall of 0.8441.
- ResNet50 improved from 78.7% to 80.6%.
- DenseNet121 improved from 78.8% to 79.7%.
- ViT showed a modest improvement from 67.0% to 69.6%.
The best-performing model (VGG16 with augmentation) surpassed the previously published AIBangla benchmark by a substantial margin.

Significance and Claims
The authors claim that their work demonstrates the efficacy of quality-aware generative augmentation for low-resource handwritten recognition tasks. By combining class-conditional diffusion synthesis with rigorous confidence-based filtering, the framework effectively addresses data scarcity and class imbalance without requiring high-resolution inputs. The results suggest that this approach offers a promising direction for improving document analysis in Indic script domains, particularly where annotated data is limited. The paper modestly notes that while the framework is computationally efficient regarding input resolution, diffusion sampling remains expensive, and the current scope is limited to isolated compound characters rather than full word recognition.