Sudoku solver using rudimentary computer vision and optical character recognition techniques. OpenCV learning project.

An image is first captured via the user's default desktop camera. The image is then taken into the script after resources are released for preprocessing.

The preprocessing step involves converting the raw image to grayscale, applying a low-pass filter (in this case, a Gaussian blur), and then applying adaptive thresholding to make the image easier to analyze.

The next phase attempts to identify the corners of the largest contour in the image and perform a perspective warp on the image. This is done by generating a kernel of ones (except the matrix corners, which are 0), performing a dilation morphological operation, finding the contours using the kernel, and then reverse-ordering the contours based on area. Filtering each contour based on length and size determines the largest one. The image is then downsampled and a top-down view of only the board is obtained.

The downsampled image is then partitioned. The warped image is resized to fit a square and is divided into 9 X 9 images, each of which represent a cell in the board (some samples below).

Using PyTesseract, each of the 81 split images is preprocessed and then analyzed via number whitelisting. Identified numbers are put into an ndarray with shape (9, 9). Once all images have been passed, the numpy array is converted into a Python list. The solver attempts to determine a solution via backtracking, which is printed to console.