Way to apply smart thresholding in images

Question

I'm writing an OCR app (for Hebrew script).

The first part in the app is thresholding,

This is how my original image looks like:

And this is how it looks like after the thresholding:

As you can see, it is mostly fine, but the "crowns" or "decorations" on the letters sometimes disappear like in this word:

That becomes:

The thing is that after I apply RGB2GRAY on the original image, the black crowns are really not dark enough, and thus they are getting white in the thresholding process, but one can see easily that it "should" be black, the question is how do I tell the algorithm to detect it...

My current thresholding code uses otzu + local thresholding, this is the code:

def apply_threshold(img, is_cropped=False):
    '''
    this function applies a threshold on the image, 
    the first is Otsu TH on all the image, and afterwards an adaptive TH,
    based on the size of the image. 
    I apply a logical OR between all the THs, becasue my assumption is that a letter will always be black,
    while the background can sometimes be black and sometimes white -
    thus I need to apply OR to have the background white.
    '''
    if len(np.unique(img)) == 2:  # img is already binary
        # return img
        gray_img = rgb2gray(img)
        _, binary_img = cv2.threshold(gray_img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        return binary_img
    gray_img = rgb2gray(img)
    _, binary_img = cv2.threshold(gray_img.astype('uint8'), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
    connectivity = 8
    output_stats = cv2.connectedComponentsWithStats(binary_img.max() - binary_img, connectivity, cv2.CV_32S)
    df = pd.DataFrame(output_stats[2], columns=['left', 'top', 'width', 'height', 'area'])[1:]
    if df['area'].max() / df['area'].sum() > 0.1 and is_cropped and False:
        binary_copy = gray_img.copy()
        gray_img_max = gray_img[np.where(output_stats[1] == df['area'].argmax())]
        TH1, _ = cv2.threshold(gray_img_max.astype('uint8'), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        # curr_img = binary_copy[np.where(output_stats[1] == df['area'].argmax())]
        binary_copy[np.where((output_stats[1] == df['area'].argmax()) & (gray_img > TH1))] = 255
        binary_copy[np.where((output_stats[1] == df['area'].argmax()) & (gray_img <= TH1))] = 0

        gray_img_not_max = gray_img[np.where(output_stats[1] != df['area'].argmax())]
        TH2, _ = cv2.threshold(gray_img_not_max.astype('uint8'), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        binary_copy[np.where((output_stats[1] != df['area'].argmax()) & (gray_img > TH2))] = 255
        binary_copy[np.where((output_stats[1] != df['area'].argmax()) & (gray_img <= TH2))] = 0
        binary_img = binary_copy.copy()
    # N = [3, 5, 7, 9, 11, 13,27, 45]  # sizes to divide the image shape in
    # N = [20,85]
    N = [3, 5, 25]
    min_dim = min(binary_img.shape)
    for n in N:
        block_size = int(min_dim / n)
        if block_size % 2 == 0:
            block_size += 1  # block_size needs to be odd
        binary_img = binary_img | cv2.adaptiveThreshold(gray_img.astype('uint8'), 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
                                                        cv2.THRESH_BINARY, block_size, 10)


    return binary_img

Any creative idea will be appreciated!

Answer 1

One approach is division normalization in Python/OpenCV.

Input:

import cv2
import numpy as np

# load image
img = cv2.imread("hebrew_text.jpg")

# convert to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# blur
blur = cv2.GaussianBlur(gray, (99,99), 0)

# divide
divide = cv2.divide(gray, blur, scale=255)

# write result to disk
cv2.imwrite("hebrew_text_division.png", divide)

# display it
#cv2.imshow("thresh", thresh)
cv2.imshow("gray", gray)
cv2.imshow("divide", divide)
cv2.waitKey(0)
cv2.destroyAllWindows()

Result:

After doing this, you may want to threshold and then clean it up by getting contours and discarding any contour with an area less than the size of your smallest accent mark.

I would also suggest, if possible, saving your images as PNG rather than JPG. JPG has lossy compression and introduces changes in color. This may be the source of some of the issues you have with extraneous marks in the background.