Odtnhj

I'm trying to clean up images containing tables of numbers for OCR.
You can see a sample here:

Test image 1

My current pipeline is as follow:

1/ Resize image to have width=256, keep aspect ratio

h, w = img.shape[:2]
ratio = 256 / w
img = cv2.resize(img, None, fx=ratio, fy=ratio, interpolation=cv2.INTER_LANCZOS4)

2/ Convert it to grayscale

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

3/ As images tend to have table borders near the edges, I remove 3px from the image borders.

gray = gray[3:-3, 3:-3]

The following 2 steps are taken from PyImageSearch

4/ Apply Gaussian blur to remove some noise

blurred = cv2.GaussianBlur(gray, (3,3), 0)

5/ Apply blackhat operator (not sure if necessary)

kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (13,5))
blackhat = cv2.morphologyEx(blurred, cv2.MORPH_BLACKHAT, kernel)

6/ Detect and remove long lines (table borders) with HoughLines

edges = imutils.auto_canny(blurred)
# horizontal lines
hlines = cv2.HoughLines(edges,1,np.pi/180,min(100,int(w*.8)),
 min_theta=np.radians(85),
 max_theta=np.radians(95))
horizontal = if hlines is None else [line[0] for line in hlines]
# vertical lines
vlines = cv2.HoughLines(edges,1,np.pi/180,min(100,int(h*.8)),
 min_theta=np.radians(-5),
 max_theta=np.radians(5))
vertical = if vlines is None else [line[0] for line in vlines]
# merge nearby lines using a long and boring function
horizontal = merge_lines(horizontal)
vertical = merge_lines(horizontal)
# draw all the remaining lines onto the blackhat image
# width=3px, color=0 (black) to remove table borders
blackhat = draw_lines(horizontal, blackhat, 0, 3)
blackhat = draw_lines(vertical, blackhat, 0, 3)

7/ (From PyImageSearch) Compute Scharr gradient, then use Otsu threshold to detect the text region

def scharr_gradient(img):
 sobel_x = cv2.Sobel(img, ddepth=cv2.CV_32F, dx=1, dy=0, ksize=-1)
 sobel_x = np.absolute(sobel_x)
 (min_, max_) = (np.min(sobel_x), np.max(sobel_x))
 sobel_x = (255 * ((sobel_x - min_) / (max_ - min_))).astype(np.uint8)
 return sobel_x

scharr = scharr_gradient(blackhat)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (20,5))
closed = cv2.morphologyEx(scharr, cv2.MORPH_CLOSE, kernel)
_, thresh = cv2.threshold(closed, 0, 255, cv2.THRESH_OTSU)

8/ Apply the mask to the original grayscale image to get a clean image

mask = np.bitwise_not(thresh).astype(np.float32)
masked = np.clip(mask + gray, 0, 255).astype(np.uint8)

The problems:

1. The Scharr operation at step 7 has a hard time detecting minus signs. Can you suggest a better way to localize the text?


2. My current pipeline fails on noisy image like this. Is there anything I can do to deal with it? I tried playing with contrast, but it worsen other cases.

Other things I tried:

1. Remove table border by detecting largest connected component didn't work because some numbers are connected to the borders, and sometimes the table borders are broken due to bad scan quality.