Pydoku is a sudoku solver in OpenCV, TensorFlow and Keras.

First clone the repository:

• With GitHub CLI:

  $ gh repo clone sRavioli/pydoku
  Cloning into 'pydoku'...

• With Git:

  $ git clone https://github.com/sRavioli/pydoku
  Cloning into 'pydoku'...

Then cd to the directory. Install the required dependencies:

pip install -r requirements.txt

or clone the conda environment:

conda env create -f pydoku.yaml

Activate the conda environment, then run the app

$ ~/.conda/envs/pydoku/python.exe ./app.py

 ┌─────────┬─────────┬────────┐
 │  2 1 5  │  3 4 7  │  9 8 6 |
 │  4 3 8  │  9 6 5  │  7 2 1 |
 │  6 9 7  │  2 8 1  │  3 4 5 |
 ├─────────┼─────────┼────────┤
 │  9 8 6  │  5 7 4  │  2 1 3 |
 │  5 7 3  │  8 1 2  │  4 6 9 |
 │  1 4 2  │  6 3 9  │  8 5 7 |
 ├─────────┼─────────┼────────┤
 │  7 5 9  │  1 2 8  │  6 3 4 |
 │  8 6 4  │  7 5 3  │  1 9 2 |
 │  3 2 1  │  4 9 6  │  5 7 8 |
 └─────────┴─────────┴────────┘

To change the sudoku image, open app.py and edit, on line 23:

# import image
original_image = ipr.read("./source/test_imgs/1.jpg")

and insert the relative path of the image.

• Upon warping and cropping an already flat image, the latter gets rotated clockwise by $90^\circ$;
• Upon determining whether a grid cell contains a number, the algorithm will sometimes consider cell containing a number empty. This is why the function sxt.construct_board() accepts an optional argument: the minimum number of white pixels;
• Upon predicting the digits, the CNN will sometimes confuse the number $1$ with the number $7$, rarely the number $6$ with number $5$ or $8$. This is probably due to the extraction process of the digit.

The content of the project itself is licensed under the GNU General Public License v3.0.