vishvam10 / clever-cell

A sudoku solver which uses OpenCV, Tensorflow and Backtracking.

CleverCell

CleverCell, a Sudoku solver, takes input as an image, solves the puzzle, and displays the result directly on the original puzzle interface.

Technologies Used

• Programming Language : Python
• Machine Learning : Tensorflow and Keras (For digit recognition)
• Libraries : Numpy (for array processing) and OpenCV (for image processing)

Basic setup

Clone the project

  git clone https://github.com/Vishvam10/AI-Sudoku-Solver.git

Go to the project directory

  cd my-project

Create a virtual environment in the project folder

  python3 -m venv /path/to/new/virtual/environment

Activate the virtual environment

  .\venv\Scripts\activate

NOTE : Please visit this link to know more about installing virtual environments

Install the dependencies using pip

  pip install - r requirements.txt

Run the program with default options. Add the -d to run the program in debug mode and -i to specify the image path

  python app.py -i "./Testing_Images/pass_1.jpg" -d

Features

• ✅ Solves sudoku of any difficulty
• ✅ Works on any type of image (.png, .jpg, etc.)
• ✅ Automatic rescaling and resizing
• ✅ Errors for unsolvable problems

Future Improvements

• One main thing would be to make it real time. The code for it is pretty straight-forward. All we need to do is pass the captured frame as the image input. Something like this :

  cap = cv2.VideoCapture(0)
  fourcc = cv2.VideoWriter_fourcc(*'XVID')
  flag = 0
  
  while(True) :
      ret, img = cap.read()
      .
      .
  
  

  While doing so, we need to proceed with the ML model only when the sudoku board is identified. And for that, a hard-coded value of image dimensions would not work. I've experimented with realtime but the results were not satisfactory (either the digits were recognized incorrectly or the board was not getting identified). There are ways to mitigate these issues and optimize the whole program. These could be implemented in the near future.

• Improving the ML model : Currently, there are a few cases where 1 or digits are misclassified, which is generally accepted to be normal. But in the case of a sudoku, even 1 misclassification can alter the problem. For example, the image located ./Testing_Images/fail_1.png fails due to this exact problem :