I've been trying for the last few days to get a sudoku grid from a picture, and I have been struggling on getting the smaller squares of the grid. I am working on the picture below. I thought processing the image with a canny filter would work fine, but it didn't and I couldn't get every contour of each square. I then put adaptive threshold, otsu, and a classic thresholding to the test, but every time, it just could not seem to capture every small square.

The final goal is to get the cells containing a number, and recognize the numbers with pytorch, so I would really like to have some clean images of the numbers, so the recognition doesn't screw up :)

Would anyone have an idea on how to achieve this? Thanks a lot in advance! :D

Here's a potential solution:

1. Obtain binary image. Convert image to grayscale and adaptive threshold

2. Filter out all numbers and noise to isolate only boxes. We filter using contour area to remove the numbers since we only want each individual cell

3. Fix grid lines. Perform morphological closing with a horizontal and vertical kernel to repair grid lines.

4. Sort each cell in top-to-bottom and left-to-right order. We organize each cell into a sequential order using imutils.contours.sort_contours() with the top-to-bottom and left-to-right parameter

Here's the initial binary image (left) and filtered out numbers + repaired grid lines + inverted image (right)

Here's a visualization of the iteration of each cell

The detected numbers in each cell

Code

import cv2
from imutils import contours
import numpy as np

# Load image, grayscale, and adaptive threshold
image = cv2.imread('1.png')
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
thresh = cv2.adaptiveThreshold(gray,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV,57,5)

# Filter out all numbers and noise to isolate only boxes
cnts = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
    area = cv2.contourArea(c)
    if area < 1000:
        cv2.drawContours(thresh, [c], -1, (0,0,0), -1)

# Fix horizontal and vertical lines
vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1,5))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, vertical_kernel, iterations=9)
horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5,1))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, horizontal_kernel, iterations=4)

# Sort by top to bottom and each row by left to right
invert = 255 - thresh
cnts = cv2.findContours(invert, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
(cnts, _) = contours.sort_contours(cnts, method="top-to-bottom")

sudoku_rows = []
row = []
for (i, c) in enumerate(cnts, 1):
    area = cv2.contourArea(c)
    if area < 50000:
        row.append(c)
        if i % 9 == 0:  
            (cnts, _) = contours.sort_contours(row, method="left-to-right")
            sudoku_rows.append(cnts)
            row = []

# Iterate through each box
for row in sudoku_rows:
    for c in row:
        mask = np.zeros(image.shape, dtype=np.uint8)
        cv2.drawContours(mask, [c], -1, (255,255,255), -1)
        result = cv2.bitwise_and(image, mask)
        result[mask==0] = 255
        cv2.imshow('result', result)
        cv2.waitKey(175)

cv2.imshow('thresh', thresh)
cv2.imshow('invert', invert)
cv2.waitKey()

Note: The sorting idea was adapted from an old previous answer in Rubrik cube solver color extraction.