datachanturia / GeoHandwritingToText

1 Symbol processing

In "./Images" folder we had handwritten texts in different formats
    We adjusted and ran "./Data-Processing/symbol_processing.py", which iterated
        over given images in "./Images/old" and extracted single characters
        from these texts
    After getting single images we manually sorted these characters and got
        "./Images/chars" folder with more than 3000 characters 300 on each
    Next step was changing image format and we manually reformated all character
        images into 25x25 format and saved them in same location

2 Data Processing

 For CNN we needed dataset, dataset in organized structure
    So we ran "./Data-Processing/data_processing.py" and all these character
        images were turned into dataset which had circa 3000 rows [elements]
        and 626 columns - 525 for each pixel from 25x25 image and 1 for
        labels - Georgian handwritten characters - which was saved in
        "./Data/train.csv"
    After getting dataset we divided it into 2 parts - training and test sets,
        training set was 90% of the data, chosen randomly from whole dataset
        and the test set was 10% of the data, which was left after randomly
        choosing training data. We saved these datasets in "./Data" folder:
        "./Data/ntrain", "./Data/ntest"

3 CNN and data training

[ Note: in this section you need "numpy", "tqdm", "matplotlib" libraries ]
- We had dataset and CNN ready and it was time to train our model
- After running "python3 train_cnn.py model.pkl" for 1-2 hours, we got our model ready
    and saved it in "./Data" folder as "./Data/model.pkl"

4 Testing the Model

- So we ran "python3 measure_performance.py" and we got our result - circa 96% accuracy.

Materials Used in Project:

https://www.coursera.org/learn/machine-learning/home/welcome
https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6
https://www.kdnuggets.com/2018/04/building-convolutional-neural-network-numpy-scratch.html
https://gluon.mxnet.io/chapter04_convolutional-neural-networks/cnn-scratch.html
https://towardsdatascience.com/convolutional-neural-networks-from-the-ground-up-c67bb41454e1
https://towardsdatascience.com/build-your-own-convolution-neural-network-in-5-mins-4217c2cf964f