BHDD using ConvNet

Update ! 3rd Jan, 2021

• Updated model architecture

• The best result - 0.99 F-score of classification is by using ConvNet with Regularization (Dropout).

• Epochs : 20

• Dropout : 0.2

• GPU Execution

• Model architecture

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 26, 26, 32)        320       
                                                                 
 max_pooling2d (MaxPooling2D  (None, 13, 13, 32)       0         
 )                                                               
                                                                 
 conv2d_1 (Conv2D)           (None, 11, 11, 64)        18496     
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 5, 5, 64)         0         
 2D)                                                             
                                                                 
 flatten (Flatten)           (None, 1600)              0         
                                                                 
 dropout (Dropout)           (None, 1600)              0         
                                                                 
 dense (Dense)               (None, 10)                16010     
                                                                 
=================================================================
Total params: 34,826
Trainable params: 34,826
Non-trainable params: 0

Burmese Handwritten Digits Dataset

• Dataset Link : https://github.com/baseresearch/BHDD

• Download data.pkl directly in order to get pkl raw file

• Train Images : 60000 with image size (28,28)

• Test Images : 27561 with image size (28,28)

• Classes : 10, i.e, handwritten digits 0 to 9

Basic ConvNet architecture

Abstract

The goal of this project is to create a model that will be able to recognize and determine the handwritten digits from its image by using the concepts of Convolution Neural Network and BHDD dataset. Though the goal is to create a model which can recognize the handwritten digits, it can be extended to letters and an individual’s handwriting. The major goal of the proposed system is understanding Convolutional Neural Network, and applying it to the Burmese handwritten recognition system.

Problem statement

To make children easier recognize digits by playing Burmese Handwritten Digits Recognizer.

Install requirements

tra@thura-pc:~$ pip install -r requirements.txt

Train BHDD with Basic ConvNet Architecture with Dropout

tra@thura-pc:~$ python cnn_train.py

• Learning Curves after training with ConvNet

Run and deploy using Streamlit

tra@thura-pc:~$ streamlit run app.py

Experiments

• We tried Single-layer Perceptron, Multi-layer Perceptron and ConvNet.
• The best result - 0.99 F-score of classification is by using ConvNet with Regularization (Dropout). Future works still need to be done for architectural innovation on BHDD.
• Epochs : 15
• GPU Execution
• Tools : OpenCV, Matplotlib, Numpy, Keras, Streamlit
• Results

User Interface Prototype

Demo

• Try it yourself on streamlit
• Demo GIF

Contributors

• Thura Aung
• Khaing Khant Min Paing
• Khant Zwe Naing

Presentations

• Project Presentation

Architecture we used

Future Works

• Need to test GANs on BHDD

References

[1] https://github.com/baseresearch/BHDD

[2] A.Dutt, A.Dutt, 2016. Handwritten Digit Recognition Using Deep Learning, International Journal of Advanced Research in Computer Engineering & Technology

[3] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. "Gradient-based learning applied to document recognition." Proceedings of the IEEE, 86(11):2278-2324, November 1998.