MyanNet V1

A Lightweight CNN for Burmese Handwritten Digit Recognition

TensorFlow • Keras • TensorFlow Lite • Edge AI

Overview

MyanNet V1 is a lightweight convolutional neural network (CNN) developed for Burmese handwritten digit recognition. The model was designed to achieve an excellent balance between recognition accuracy and computational efficiency, making it suitable for deployment on mobile, embedded, and resource-constrained devices.

Instead of maximizing accuracy through increasingly deeper networks, MyanNet V1 focuses on reducing computational complexity while maintaining competitive performance. The architecture combines depthwise separable convolutions, batch normalization, global average pooling, and dropout regularization to produce a compact yet highly effective classifier.

The model was trained and evaluated on the Burmese Handwritten Digit Dataset (BHDD), containing 87,561 handwritten digit images across ten Burmese numeral classes.

Model Highlights

Property	Value
Task	Burmese Handwritten Digit Recognition
Framework	TensorFlow / Keras
Input Size	28 × 28 Grayscale
Classes	10 (၀–၉)
Trainable Parameters	10,634
Test Accuracy	99.49%
5-Fold CV Accuracy	99.46% ± 0.06%
Quantized Model Size	24.18 KB
Average CPU Inference	0.263 ms/image
License	Apache-2.0

Architecture

MyanNet V1 follows a compact two-stage convolutional architecture optimized for lightweight deployment.

Feature Extractor

Standard Convolution
Batch Normalization
Max Pooling
Depthwise Separable Convolution
Batch Normalization
Max Pooling

Classification Head

Global Average Pooling
Dropout
Fully Connected Layer
Softmax Output

The architecture significantly reduces parameter count while preserving classification accuracy.

Performance Comparison

MyanNet V1 achieves competitive accuracy while reducing the number of trainable parameters by approximately 69.5% compared to the baseline CNN.

Model	Parameters	Accuracy
Baseline CNN	34,826	99.58%
GAP-BN CNN	21,418	99.51%
MyanNet V1	10,634	99.49%

Training Curves

The training and validation curves demonstrate stable convergence with minimal overfitting.

Confusion Matrix

The confusion matrix shows strong classification performance across all Burmese digit classes.

5-Fold Cross Validation

The model was evaluated using stratified five-fold cross validation to assess robustness and generalization.

Metric	Value
Mean Accuracy	99.46%
Standard Deviation	0.06%

Sample Predictions

Example handwritten digit samples from the BHDD dataset.

Failure Cases

Representative misclassified samples produced by MyanNet V1.

Although misclassifications are rare, they primarily occur for ambiguous handwriting styles and visually similar digit shapes.

Dataset

The model was trained and evaluated on the Burmese Handwritten Digit Dataset (BHDD).

Dataset Summary:

87,561 handwritten digit images
10 Burmese digit classes (၀–၉)
28 × 28 grayscale images
MNIST-compatible format

Training Split:

60,000 balanced training images

Testing Split:

27,561 naturally imbalanced testing images

Please obtain the dataset from the official BHDD repository.

Usage

TensorFlow / Keras

import tensorflow as tf

model = tf.keras.models.load_model("myannet_best.keras")

prediction = model.predict(image)

TensorFlow Lite

import tensorflow as tf
import numpy as np

interpreter = tf.lite.Interpreter(
    model_path="myannet_quantized.tflite"
)

interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

image = np.expand_dims(image / 255.0, axis=(0, -1)).astype(np.float32)

interpreter.set_tensor(input_details[0]["index"], image)

interpreter.invoke()

prediction = interpreter.get_tensor(output_details[0]["index"])

Intended Applications

MyanNet V1 is suitable for:

Burmese handwriting recognition
Educational software
Mobile OCR systems
Embedded AI devices
Edge computing
Research benchmarking

Limitations

MyanNet V1 was trained exclusively on isolated handwritten Burmese digits contained in the BHDD dataset.

The model has not been evaluated on:

Printed Burmese text
Burmese characters
Burmese words or sentences
Historical manuscripts
Camera-captured documents
General OCR tasks

Performance outside the BHDD domain may differ significantly.

Version Information

This repository contains the original public release of MyanNet (Version 1).

Version 1 established the lightweight CNN architecture and serves as the baseline for future iterations.

Future versions aim to improve:

Model compactness
Computational efficiency
Robustness
Recognition performance
Support for larger Burmese OCR tasks

Citation

If you use MyanNet V1 in your research, please cite:

@software{maung2026myannetv1,
  author    = {Ah Maung Oo},
  title     = {MyanNet V1: A Lightweight CNN for Burmese Handwritten Digit Recognition},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Drew2456/MyanNet-V1}
}

Once the associated journal paper is published, this citation will be updated with the official publication.

Acknowledgements

BHDD dataset by Swan Htet Aung et al.
TensorFlow & Keras
Optuna for hyperparameter optimization
The open-source machine learning community

⭐ If you find MyanNet V1 useful, please consider starring the GitHub repository and citing this work in your research.

Downloads last month: 9

Drew2456
/

MyanNet-V1