It's includes following steps:
- Data collection and Datasets Generation
- Train and Testing model
- We have visited more than 30 schools with in the Pokhara valley.
- About 2000 students are involved directly during dataset collection.
- We provided students empty A4 paper and they write down characters accordingly.
- All A4 paper are scanned and converted to .tif image form.
- Image are pre-processed and characters are generated in following approach: Python implementation here
- Data preprocessing: We already have preprocesse data during dataset generation.
- Convolutional Neural Network is used to train our dataset. Notebook-1here
- About 99% training accuracy and 93% validation accuracy has been obtained.
- Model is overfitted,so I applied BatchNormalization and Dropout. But result is not promising.Notebook 2 here
- Further works need to be done in the area of hyperparameter tuning since the model is overfitted.
- Word and line recognition will be further research area.