Burmese Handwritten Digit Dataset (BHDD)
- BHDD is a standardized dataset featuring 87,561 grayscale images of handwritten Burmese digits in MNIST format, serving as a benchmark for OCR research.
- It offers diverse samples from over 150 contributors with balanced training and unbalanced test splits to reflect realistic handwriting variability.
- Baseline models including MLP and CNN achieve high accuracy, while structured errors highlight intrinsic challenges in recognizing rounded Burmese script.
The Burmese Handwritten Digit Dataset (BHDD) is a collection of 87,561 grayscale images of handwritten Burmese digits in ten classes, with each image standardized to pixels in the MNIST format. It was introduced to provide a public benchmark for Burmese handwritten digit recognition, a setting for which the dataset paper states there had been no public benchmark despite the centrality of MNIST-like resources in Latin-script OCR. A later benchmark paper explicitly states that myMNIST was formerly called BHDD and that myMNIST and BHDD refer to the same dataset, with the newer name emphasizing its role as a Burmese counterpart to MNIST (Aung et al., 23 Mar 2026, Thu et al., 19 Mar 2026).
1. Naming, scope, and motivation
BHDD stands for Burmese Handwritten Digit Dataset. The dataset was created because, although MNIST and similar datasets established a standard for Latin-script digit recognition, there was no public benchmark for Burmese handwritten digits. The dataset paper argues that Myanmar script is visually distinct because it is traditionally written with rounded forms, described as “sar-lone,” or “round script,” and that these shapes produce classification challenges that differ from Latin digits (Aung et al., 23 Mar 2026).
The benchmark paper adopts myMNIST as the dataset name used in that study and states explicitly that it was formerly called BHDD. It is equally explicit that myMNIST = BHDD, that BHDD is a publicly available Burmese handwritten digit dataset, and that the resource is intended to serve as the direct Burmese-language analogue of MNIST (Thu et al., 19 Mar 2026).
A common source of confusion is therefore nomenclature rather than corpus content: in the documentary record provided by the two papers, BHDD and myMNIST refer to the same dataset, not to different releases or different corpora. The significance of the resource lies in its use as a standardized benchmark for Burmese OCR and handwritten digit recognition, with an emphasis on script-specific ambiguity produced by rounded, loop-based glyph morphology.
2. Dataset composition and split design
BHDD contains 87,561 grayscale images of handwritten Burmese digits across 10 classes labeled from 0 to 9. Every image is 28 × 28 pixels, matching the MNIST format, and each sample is stored as an 8-bit grayscale image with values from 0 to 255. The benchmark paper also describes the images as represented in model-ready form as flattened vectors of length 784, with labels provided in one-hot encoded format (Aung et al., 23 Mar 2026, Thu et al., 19 Mar 2026).
| Split or property | Value | Notes |
|---|---|---|
| Total images | 87,561 | 10 classes |
| Training set | 60,000 | Balanced, 6,000 per class |
| Test set | 27,561 | Unbalanced, natural collection frequencies |
| Image size | MNIST format | |
| Pixel format | 8-bit grayscale | Values from 0 to 255 |
The training set is explicitly balanced, with 6,000 samples per class. The test set is intentionally unbalanced, reflecting the natural frequencies that arose during collection; the dataset paper reports that per-class counts range from 6,856 images for class 0 down to 389 images for class 9 (Aung et al., 23 Mar 2026).
The split protocol is one of the dataset’s most consequential design choices. The training and test splits are made by contributor, so no writer’s handwriting appears in both sets, which helps prevent identity leakage. The dataset is also verified to contain no exact duplicates within or across splits (Aung et al., 23 Mar 2026). For benchmark construction, this means that reported performance is less likely to depend on writer-specific memorization and more likely to reflect actual class discrimination.
3. Collection process, extraction pipeline, and quality control
The collection process was community-based and large-scale. The authors organized the effort through the Expa.AI Research Team, and over 150 people contributed samples. Contributors wrote digits on plain A4 paper, both white and yellow sheets, typically filling ten or more pages each, with roughly 500–600 digits per page. In total, about 2,500 sheets were collected (Aung et al., 23 Mar 2026).
Contributors came from several groups, including volunteers and interns from Taungoo Computer University, high school students from St. Augustine / B.E.H.S (2) Kamayut, and friends and family of the research team. The age range spanned from teenagers to people in their 50s, and reported occupations included students, clerks, and programmers. Most contributors were from Yangon, with smaller groups from Mandalay, Nay Pyi Taw, Shan State, and the United States (Aung et al., 23 Mar 2026). The benchmark paper summarizes this diversity more compactly, describing the dataset as curated from over 150 individuals spanning a wide range of ages, from high school students to professionals in their 50s, and spanning different occupations (Thu et al., 19 Mar 2026).
A practical collection constraint was that most sheets were photographed with phone cameras rather than scanned, so lighting, perspective, and camera quality varied substantially. To handle this, the team built an Android application that performed adaptive thresholding and contour detection in real time, allowing contributors to inspect whether the digits were extracted cleanly before submission. The app also saved preprocessing parameters as metadata (Aung et al., 23 Mar 2026).
The extraction pipeline was based on OpenCV and proceeded by converting images to grayscale, applying adaptive thresholding, finding contours to isolate each digit, cropping each digit to its bounding box, centering it in a 28 × 28 frame, and normalizing pixel intensity. Quality control was performed by a 20-member annotation team using a chatbot-based review tool, followed by a final verification pass by two data engineers. Mislabeled, illegible, and duplicate images were removed at each stage (Aung et al., 23 Mar 2026). In methodological terms, the dataset is therefore not merely a scanned handwriting archive but a curated recognition benchmark with explicit extraction and filtering stages.
4. Visual statistics, morphological variation, and confusion structure
The dataset paper provides several analyses of BHDD’s visual and statistical properties. The mean pixel intensity per class ranges from 12.5 for class 2 to 26.5 for class 0. The ink coverage—the fraction of non-zero pixels—ranges from 30.4% for class 2, described as a thin hook, to 56.8% for class 0, described as a full circle (Aung et al., 23 Mar 2026).
Mean images computed over the 6,000 training samples per class show that each digit has a distinct canonical shape. The paper gives specific examples: class 0 is a ring, class 1 is an open arc, and class 8 is a spiral. It emphasizes that these averages are relatively sharp, which suggests writers agree on the broad form of each digit despite individual variation (Aung et al., 23 Mar 2026).
Morphological variation was examined using per-class variance heatmaps. The largest variance occurs at stroke endpoints and junctions, where handwriting styles diverge most. More complex digits, such as class 8, have broader regions of high variance, while simpler digits like class 2 vary in a narrower band. The paper also highlights within-class stylistic diversity in classes 0, 3, 5, and 8, where differences in stroke thickness, curvature, and loop closure are especially visible (Aung et al., 23 Mar 2026).
The confusion structure is one of the most distinctive aspects of BHDD. The most difficult pair is 0 and 1, which are confused with each other 24 times in total; the main difference is whether the circle is closed or has a small gap. Another problematic pair is 0 and 8, which share a round outer form but differ internally, leading to 8 combined errors. Digit 3 is mistaken for 1 seven times when its tail is faint, and 5 and 8 sometimes overlap as well, with 3 errors (Aung et al., 23 Mar 2026). The benchmark paper likewise presents confusion analysis as evidence of dataset difficulty, highlighting 0 vs 1, 1 vs 5, and 3 vs 4, and attributing these confusions to the intrinsic ambiguity of handwritten Burmese glyphs (Thu et al., 19 Mar 2026).
These analyses situate BHDD as a script-specific recognition benchmark rather than a simple transposition of the MNIST template. The dataset paper’s conclusion is that when digits share curved sub-strokes, small variations in pen pressure or loop closure can make one class look like another (Aung et al., 23 Mar 2026).
5. Baseline models and benchmark results
The dataset paper reports three baseline models, all trained with seed 42: an MLP, a CNN, and an Improved CNN. The MLP is a multilayer perceptron with two hidden layers of 256 and 128 units, ReLU activations, Adam optimization, and early stopping on a 10% validation split, trained on flattened 784-dimensional inputs. The CNN is a two-layer convolutional network with 32 and 64 filters of size 3 × 3, ReLU activations, 2 × 2 max pooling, dropout of 0.25 in convolutional layers and 0.5 in the dense layer, followed by a 128-unit fully connected layer and a 10-way output layer; it was trained for 15 epochs with Adam at learning rate using cross-entropy loss. The Improved CNN is a deeper network with three convolutional layers using 32, 32, and 64 filters, batch normalization after each convolution, two max-pooling stages, and a 128-unit fully connected layer; it has 431K parameters, only slightly more than the baseline CNN’s 421K, and was trained for 25 epochs with on-the-fly augmentation including rotation ±15°, translation ±2 px, and scale 0.9–1.1×, plus cosine learning-rate annealing and light weight decay (Aung et al., 23 Mar 2026).
| Model | Architecture summary | Reported test results |
|---|---|---|
| MLP | 256/128 hidden units | Accuracy 99.40%, Macro F1 0.9905 |
| CNN | 2 conv layers, 32/64 filters | Accuracy 99.75%, Macro F1 0.9964 |
| Improved CNN | 3 conv layers, batch norm, augmentation | Accuracy 99.83%, Macro F1 0.9980 |
The corresponding macro precision and macro recall are also reported. For the MLP they are 0.9876 and 0.9934; for the CNN they are 0.9959 and 0.9970; and for the Improved CNN they are 0.9972 and 0.9988 (Aung et al., 23 Mar 2026). The Improved CNN’s error analysis is especially compact: out of 27,561 test samples, only 47 are misclassified, and about half of those errors involve the 0–1 pair, reinforcing the claim that this pair is structurally the hardest to separate.
A separate benchmark paper evaluates eleven architectures on the same dataset under a unified PyTorch training pipeline: MLP, CNN, LSTM, GRU, Transformer, JEM, FastKAN, EfficientKAN, PETNN (Sigmoid), PETNN (GELU), and PETNN (SiLU). The training setup is reported as PyTorch, a single NVIDIA GeForce RTX 3090 Ti GPU (24 GB VRAM), AdamW, learning rate , weight decay , OneCycleLR, max LR , batch size 128 for training and 1000 for validation/test, cross-entropy loss, dropout, layer normalization, gradient clipping with max norm 1.0, 30 to 100 epochs depending on model, early stopping based on validation loss plateau, and random seed 42 (Thu et al., 19 Mar 2026).
Under that benchmark, CNN is reported as best overall with Precision 0.9955, Recall 0.9963, F1-Score 0.9959, and Accuracy 0.9970. PETNN (GELU) follows with Precision 0.9947, Recall 0.9963, F1-Score 0.9955, and Accuracy 0.9966, while PETNN (SiLU) reaches F1-Score 0.9952 and Accuracy 0.9964. JEM is also competitive with F1-Score 0.9944 and Accuracy 0.9958. The benchmark paper describes FastKAN and EfficientKAN as meaningful alternative baselines, with accuracy roughly around 0.992, while MLP is the weakest among the reported models but still reaches Accuracy = 0.9907 (Thu et al., 19 Mar 2026).
Taken together, these results establish BHDD/myMNIST as a high-accuracy but nontrivial benchmark: strong convolutional models approach saturation, yet the residual errors remain structured and script-specific rather than random.
6. Availability, downstream relevance, and documentary ambiguities
The dataset paper states that BHDD is released at https://github.com/baseresearch/BHDD under CC BY-SA 4.0. The repository includes pickle and gzip-compressed IDX formats, with the IDX files designed to work directly with standard MNIST data loaders. It also provides exploration scripts, baseline code, and usage examples (Aung et al., 23 Mar 2026).
The benchmark paper also cites the repository as Expa.AI Research Team, “Burmese Handwritten Digit Dataset (BHDD)”, available at the same URL, but states that the dataset is publicly available under the LGPL-3.0 license and intended to foster benchmarking and innovation for Burmese AI research (Thu et al., 19 Mar 2026). The two papers are therefore aligned on public availability and repository location, but not on license wording. This suggests that researchers should verify the current repository state and citation practice directly at the project repository.
In the benchmark paper, the dataset’s broader relevance is framed in relation to Myanmar NLP/AI research, with the argument that Myanmar/Burmese digits have distinctive morphology, especially looped strokes and curved ligatures, and that these shapes create systematic confusions not present in Latin-script digit recognition. The paper identifies a Burmese digit benchmark as relevant to e-government, finance, education, archival digitization, and the broader Myanmar NLP/AI ecosystem (Thu et al., 19 Mar 2026).
BHDD’s importance therefore lies in three linked properties: it is a contributor-diverse corpus; it is standardized to the MNIST interface while preserving script-specific visual ambiguity; and it is accompanied by explicit baseline and benchmark results. Within the current literature, it functions both as a dataset for Burmese handwritten digit recognition and as a reproducible testbed for comparing classical deep learning, recurrent, transformer, KAN, energy-based, and PETNN approaches on a regional script with rounded handwritten forms (Aung et al., 23 Mar 2026, Thu et al., 19 Mar 2026).