分割图像中的连接字符

Question

My aim is to correctly segment the characters in an image

My image looks like this:

How can I correctly segment the connected B and W? Also my code seems to find that 750 are all connected as well. How do I segment them?

Which transformation must I apply? I tried erode but it did not help? How is the kernel size selected for such an image? How should I remove the noise on 5 and M?

What changes should I make to my code to correctly segment and isolate every character? Code:

img = cv2.imread('C:\\xx\\testimages\\X\\plate4.jpg', 0)
cv2.threshold(img,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU,img)
image, contours, hier = cv2.findContours(img, cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_NONE)
contours = sorted(contours, key=lambda ctr: cv2.boundingRect(ctr)[0])
d=0
for ctr in contours:
    # Get bounding box
        x, y, w, h = cv2.boundingRect(ctr)
    # Getting ROI
        if w>20 and h>20 and w<60:       #Boundary conditions to isolate a character
            print(x, y, w, h)
            roi = image[y:y+h, x:x+w]
            #roi=cv2.resize(roi,(20,35))
            #kernel = np.ones((3,3), np.uint8)
            #roi = cv2.morphologyEx(roi, cv2.MORPH_CLOSE, kernel)
            #roi = cv2.erode(roi, kernel, iterations=1)
            # kernel_1=np.ones((1,1),np.uint8)
            # roi=cv2.dilate(roi,kernel,iterations=1)
            cv2.imshow('character: %d' % d, roi)
            cv2.imwrite('C:\\xx\\ValidationSet\\character_%d.png'%d, roi)
            cv2.waitKey(0)
            cv2.destroyAllWindows()
            d+=1

Answer 1

As people recommended in the comments already, the best options is to use morphological transforms, like erosion and opening. Regarding the kernel size, you can make it a function of the width of your contours, or just iterate multiple times applying a small kernel of size (3, 3) or (5,5). I personally found this tutorial (in C++, but the concept is the same) simple and useful. Regarding noise removal from the "5" and "M", you can't do that in a generic way that will generalize to all possible types of noise. If you have more information on the statistics of the errors or some more information (eg the "noise" always manifest in the form of white pixels on top of the current character) it is easier and doable, of course.