bounding boxes not found

[kubasiak]

Hi, First of all thank you for this repo. I enjoyed it a lot. I see that there is an issue with the boxes around characters. When using cv2.findContours sometimes the contours are not found. I find that it is in large extent because of the frames that are around the plates. I have solved it with removing a margin of 5% from the plate picture. ( Later of course the coordinates of the frames must be adjusted for this margin in respect to the original picture.)

It is not the most elegant solution but it helped:

    cont, _  = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    marginh=0
    marginw=0
    if(len(cont)<=1):
        print('havent found any characters -  cutting margings')
        m=0.05
        marginh=round(m*maxh)
        marginw=round(m*maxw)
        cut_plate_image=plate_image[marginh:maxh-marginh,marginw:maxw-marginw]
        cut_binary=binary[marginh:maxh-marginh,marginw:maxw-marginw]
        cont, _  = cv2.findContours(cut_binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        test_roi = cut_plate_image.copy()

    cut_plate_image=plate_image[marginh:maxh-marginh,marginw:maxw-marginw]
    cut_binary=binary[marginh:maxh-marginh,marginw:maxw-marginw]

And then of course adjust when cropping images:

# extract the img
            curr_num = cut_binary[y:y+h,x:x+w]
            curr_num = cv2.resize(curr_num, dsize=(digit_w, digit_h))
            _, curr_num = cv2.threshold(curr_num, 220, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
            curr_num_orig = cut_plate_image[y:y+h,x:x+w]
            crop_characters.append(curr_num) 
            crop_characters_orig.append(curr_num_orig)

[modevera]

I have the same issue how didi you define maxh ?

[eaglemas]

Instead of : ,cont, = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

write : ,cont, = cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

and everything is fixed!!!!

[atharva21jadhav]

@eaglemas I'm working with RETR_TREE too currently. Although it can work okay at such scenarios, RETR_TREE might also create extra contours. For example, for digits like 0 since all the contours hierarchy is calculated we might get 2 contours one inside and the other on the external. It would be better if we go with RETR_EXTERNAL and crop the portions of image.

[amandp13]

@kubasiak Can you please provide me the optimal code for this issue as I am getting confused by your soluton?