Custom Handwritten Character Recognition > Go to project

Aim

This project was designed to use the EMNIST Dataset to create a custom model for classifying the following target characters, as well as artificially constructed characters with a strikethrough:

• TARGET_CLASSES = [‘0’,’1’,’2’,’3’,’4’,’5’,’C’,’M’,’V’,’X’]
• STRIKETHROUGH_CLASSES = [‘0_s’,’1_s’,’2_s’,’3_s’,’4_s’,’5_s’,’V_s’]

Data

The EMNIST data can be found at EMNIST and was imported the TFDS API.

Process

• Loaded data through the TFDS API using builder.as_data_source() as data is in ArrayDataSource format.
• Filtered data by Target Classes and creates artificially constructure Strikethrough Classes.
• Prepared Data for training including normalizing images and one hot encoding labels.
• Viewed samples to verify clean data.
• Create a baseline model and performed a series of experiments, including Batch Normalization, Dropout, and adding complexity to avoids overfitting.
• Used Data Augmentation on final model to mimic real world data.
• Produced a final model based on experiments and tested on data.

Conclusion

We successfully created a model which predicted a class with 99.5% accuracy.