mythrex / omr-scanner Goto Github PK

Due to uneven lighting, the Otsu thresholding is not working properly.
The reason, I don't know.

The bubble threshold is creating a problem. A bubble_thresh for one is not working for other. Take a look at images. The threshold is working for test-2 but not for test-3.

# sort the question contours from top to bottom
questionCnts = contours.sort_contours(questionCnts, method='top-to-bottom')[0]
correct = 0
# sort the first 4 countours left to right
e = contours.sort_contours(questionCnts[0:4])[0]
# this should draw 1 a countour
cv2.drawContours(paper, [e[0]], -1, 255, -1)

This code should show question 1 A contour but instead it shows some random contour.
See in pic which contour is being displayed.

# sort the question contours from top to bottom
questionCnts = contours.sort_contours(questionCnts, method='top-to-bottom')[0]
correct = 0
# sort the first 4 countours left to right
e = contours.sort_contours(questionCnts[0:4])[0]
# this should draw 1 b countour
cv2.drawContours(paper, [e[1]], -1, 255, -1)

This code should show question 1 B contour but instead it shows some random contour.
but instead the following has been showing.

for c in filter_cnts:
    (x, y, w, h) = cv2.boundingRect(c)
    ar = w / float(h)
    if w >= 9 and h >= 9 and ar >= 0.7 and ar <= 1.2:
       # change this!!!!!!!!!!!!!!!!1
        box = [(x//5)*5, round(y/20.0)]
        questions.append([c, box])
        cv2.rectangle(paper, (x, y), (x+w, y+h), (0, 255, 0), 1)

Problem is with calculation of box for sorting in top to bottom.
Change this logic.
box = [(x//5)*5, round(y/20.0)]

As you can see in this screenshot the system is detecting non bubbled answers.

The box-line for bubbles is too thick. The system sometimes takes it as two contours and the most external contour is the box because of which RETR_EXTERNAL return method fails.

The code is monolithic. We need to modularize it.

We have a use case where we need to identify radio button group.
radio button group can have 'n' number of radio button on the same line or multi lines.
Will this solution work?

The module should throw an error in some cases like

• when it does not detect all the bubbles
• if the paper is not detected
• if only external contour after the perspective transform is detected.

As you can see in screenshot a lot of insignificant contours are detected but the main contour is not detected.

I am very interested in this project. However I don't know how to use it exactly, could hints about installation be written in the readme ?

Refer this pic for more details.

Changing the min height and width is creating a problem. D is being detected.

if w >= 9 and h >= 9 and ar >= 0.7 and ar <= 1.2:

Look at image for more details.

The external return method for findContours is not working.

Because it is only finding a contour with an external box and all the contours inside it are not detected.