Sudoku will be extracted from image and digits to be recognized for further solving sudoku
- Trained a CNN model on MNIST dataset for digit recognizer and achived an accuracy of 97.87 % .
- Below steps taken for extracting sudoku from a given image :
- Computed a threshold mask image based on local pixel neighborhood and extracted image above the threshold value to avoid any noise present in image.
- Extracted contours from image and sorted them in descending order.
- Traversed contour and approximated a polygonal with the specified tolerance to find exact coordinates of sudoku grids.
- Coordinates received was transformed over 270 x 270 matrix.
- Removed borders of 1's or 0's.
- Cropped images into 30 x 30 shape.
- Cropped images were used detect digit after border removal.
- Boxes which were blank and needs to be predicated were given 0 as digit.
- Used backtracking to solve sudoku.
CNN Model used for training
- nn.Conv2d(1,20,kernel_size=5)
- max_pool2d(,2)
- relu
- nn.Conv2d(20,40,kernel_size=5)
- nn.Dropout2d()
- max_pool2d(,2)
- relu
- nn.Linear(640,120)
- dropout
- nn.Linear(120,10)
- log_softmax
Image after masking and finding grids coordinates
Image after suduko grid was cropped out from orignal image
