Sudoku Solver

Sudoku solver app using OpenCV for image processing and Pytorch for training model on custom printed digits dataset.
Finished time: 12/01/2023.

About dataset and model

We don't use MNIST dataset because:

+) A large number of digits from input images are printed digits (digital digits).

+) There are some samples which have wrong labels (noise), the model for sudoku digits classifier is need to be as accurate as possible.

+) There are some mistakes recognizing digits (especially 0, 1 and 7, 3 and 8) when using MNIST dataset.
In this repo, we create a custom printed digit dataset by drawing digits with various popular font styles (Times New Roman, Tahoma, Arial, Hevectica ...) which are shown in font folder.

Loading custom dataset in Pytorch

A custom Dataset class must implement three functions: __init__, __len__, and __getitem__.
Tutorial link: https://pytorch.org/tutorials/beginner/basics/data_tutorial.html.
PrintedMNIST class for custom dataset with Pytorch:

 class PrintedMNIST(Dataset):
    """
    Generate digital mnist dataset for digits recognition
    """
    def __init__(self, samples, random_state, transform = None):
        self.samples = samples
        self.random_state = random_state
        self.transfrom = transform
        self.fonts = fonts
        random.seed(random_state)
        
    def __len__(self):
        return self.samples
        
    def __getitem__(self, index):
        color = random.randint(200,255)
        
        #Generate image
        img = Image.new("L",(256, 256))
        label = random.randint(0,9)
        size = random.randint(180, 220)
        x = random.randint(60, 80)
        y = random.randint(30, 60)
        draw = ImageDraw.Draw(img)
        
        #Choose random font style in font style list
        font = ImageFont.truetype(random.choice(self.fonts), size)
        draw.text((x,y), str(label), color, font = fonts)
        img = img.resize((28,28), Image.BILINEAR)
        if self.transfrom:
            img = self.transfrom(img)
        return img, label

Model: Pretrained ResNet50, adjust the first convolutional layer to feed gray-scale image after image processing step.

Overall Image Processing Pipeline

Original Image	Adaptive Threshold

Find corners	Perspective Transform

Find grid mask	Extract digit images

Clean noises	Recognize digits

Solve board	Draw solved board on image

Inverse Perspective Transform	Final result

Note

Change the url variable to your image url.

Test sudoku solver by image locally, run image_solver.py.
Test real-time solver locally, run realtime_solver.py.
Test the entire image processing pipeline, run test_modules.py.

Sudoku Solver App

Streamlit is a free and open-source framework to rapidly build and share beautiful machine learning and data science web apps. It is a Python-based library specifically designed for machine learning engineers.
To run app, type the following line on cmd:
```
streamlit run app.py
```
The Web application has 6 pages:
- Home page
- Sudoku solver by number inputs.
- Sudoku solver by image.
- Real-time sudoku solver.
- About Sudoku.
- About me.

Home Page

Sudoku solver by number inputs page

Surface.

Get input sudoku puzzle from user.

Empty puzzle	User input

Generate input image and solved image from input puzzle.

Input image	Result image

Sudoku solver by image inputs page

Surface.

Upload image and get result image.

Uploaded image	Result image

Real-time sudoku solver page

Surface.

About Sudoku page

Some information and rules of this game.

Rules	How to play

About me page

Just some personal information.

Comments

Pros

Good performence of recognizing digits task.
Acceptable response time for most of input images upload to the App.
Simple dataset and don't need much data to get high accuracy.
Clear overall structure and pipeline, there are test modules for testing and debugging.
Recognize and solve correctly most of input images with large shape, high resolution and taken form bird's eye view.
Perspective transformation works effectively with an acceptable proportion of the input images.
There are exception processes when receiving invalid input from user and when there is something wrong with recognize digit task or solve Sudoku task.

Cons

Real-time solver is not really effective in some cases.
Face problems setting threshold value to clean noises of each squares, sometimes cleaning noise also cleans the digit contained in that square.
It takes long time to solve hard level Sudoku board (>10s).
Wrongly recognize and solve small shape, unclean and blurred images or images with too inclined perspective toward Sudoku board.

Development orientation

Get more robust model for digits recognizing task.
Apply Denoising Autoencoders for cleaning noise task to handle the problem of setting threshold value.
Improve real-time solver function.
Get better web app interface design.