Status: I haven't touched this project since the course ended. Even though the method seems promising, it currently has a overfitting problem because the generator generates too simple training data. So if you want to improve upon this, start from improving the training data generator.

Dependencies are managed with pipenv.

The first you need is a text file of Russian text. This model was trained with this Russian News Corpus, which you can get from here. You can also download the generated training data from here

The code to compare against Tesseract is located in test directory. To run it, you need to install Tesseract 4 properly. See this for more information.

A pretrained model can be downloaded from Google Drive

The model can be trained at Google Colaboratory.

Once you have installed the pipenv dependencies, you can activate the virtualenv with:

pipenv shell

To make a prediction on your own machine, run predict.py like following:

python predict.py --model <path/to/model> predict_this.png

NOTE: The input must be a grayscale image with the height of 32 pixels. While compare_tesseract.py and the training cannot handle JPEG files, they should work fine with predict.py.

To generate training data yourself

To run automated comparison against Tesseract, go to test directory, and run the following:

python compare_tesseract.py -m <path/to/model> <input_directory>

The content of the directory can be generated with

Each image in the input_directory must be a PNG file, and must have a corresponding .gt.txt file, which contains the ground truth. For example, if there exists an image with the name 1.png, there must also be a file 1.gt.txt.

Apache 2.0