The dataset is downloaded from Akshar. (https://github.com/harshthaker/Akshar). This dataset Akshar is a free, open-source dataset containing 3400 images of Gujarati language characters. The dataset includes a wide variety of distinct characters because of different negotiations in writing the same character. The characters are made available for download as TIFF files. Some of these character images are very complex shaped and closely correlated with others.
The dataset used in the project is preprocessed dataset. The algorithm used for preprocessing is also included with the name train. ipynb. All the characters in the dataset were not used as some of them were similar images with different labels. I used 34 unique characters, and each character is represented in 100 unique ways or notations.
Software Requirements:
- TensorFlow 1.2.1
- Keras
- matplotlib
- scikit-learn
Future Scope
- Use multiple Algorithms
- Add some more languages
- Optimize CNN Model
- Configure model to Web and mobile applications