srdg / bangla-dl

Classifying handwritten Bengali alphabet characters in real-time using OpenCV and Keras.

bangla-dl

Classifying handwritten Bengali alphabet characters in real-time.
Implementation of Handwritten Bangla Digit Recognition using Deep Learning by Md Zahangir Alom et. al.

Dataset

The CMATERdb pattern recognition database.

Dependencies

Clone the repository and move into the src folder. Run in terminal : pip3 -r install requirements.txt

Architecture

Conv2D - MaxPool - Conv2D - MaxPool - DropOut - FC - Softmax(Classification) for custom model
RESNET50 for transfer learning

Accuracy

Train : ~86%
Test : 86% 91%

Sample recognition recorded in real-time

Remarks

• Image size used is 32 x 32
• The model is still not robust enough, accuracy can be improved using inception.

References

• Inspired from the work of @akshaybahadur21.
• Help taken from
  • OpenCV official documentation
  • Multi-label classification with keras
• Looking forward to improve accuracy from this implementation of @dibyatanoy.

To-do

• Use transfer learning on multiple models like VGG, DenseNet, SqueezeNet
• Predict in real-time using fast.ai
• Update real-time prediction to be done with tracking instead of handwritten characters.