Separate text from commentary automatically in an image using OpenCV

autocrop.py

import cv2
import glob
import os
import math
import numpy as np

# Parameters:
# xgive is the number of pixels of "give" when determining whether a line of text
# belongs to the main text or commentary, when they are discriminated based on
# width.
xgive = 70

images = [(cv2.imread(file),os.path.basename(file)) for file in glob.glob("*.tif")]

for image in images:
    img = image[0]
    newimg = img.copy()
    height, width, channels = img.shape
    filename = image[1]
    out = "bounding_boxes/" + os.path.splitext(filename)[0]+"-new.tif"
    roiname = "cropped/" + os.path.splitext(filename)[0]+"-roi.tif"

    # Dilate and erode the inverted image to get horizontal lines
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    blur = cv2.GaussianBlur(gray,(15,15),0)
    thresh = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]
    kernel = np.ones((1,20), np.uint8)  # note this is a horizontal kernel
    d_im = cv2.dilate(thresh, kernel, iterations=6)
    e_im = cv2.erode(d_im, kernel, iterations=6) 
    new_d_im = cv2.dilate(e_im, kernel, iterations=2)

    # Now find the contours (bounding boxes)
    contours, hierarchy = cv2.findContours(new_d_im, cv2.RETR_EXTERNAL, 
                                                cv2.CHAIN_APPROX_NONE) 
    
    # First, get the maximum width of the bounding boxes.
    maxwidth = 0
    for cnt in contours: 
        x, y, w, h = cv2.boundingRect(cnt) 
        if w > maxwidth:
            maxwidth = w

    # Now, to find the first line in the commentary, we see whether
    # a given bounding box is equal to, or a few pixels less than,
    # the maximum width of the bounding box.
    wideboxes = []
    yvalues = []
    for cnt in contours:
        x, y, w, h = cv2.boundingRect(cnt) 
        if (w >= (maxwidth - xgive)):
            wideboxes.append(cnt)
            yvalues.append(y)

    # now, we toss out bounding boxes that are more than 
    # 1 standard deviation 
    # away from the nearest other bounding box to get rid of outliers
    filtered = []
    mean = np.mean(yvalues, axis=0)
    sd = np.std(yvalues, axis=0)
    for cnt in wideboxes:
        x, y, w, h = cv2.boundingRect(cnt) 
        if not (y < sd):
            filtered.append(cnt)

    # From the FILTERED list, we get the maximum Y value.
    maxy = img.shape[0]
    for cnt in filtered:
        x, y, w, h = cv2.boundingRect(cnt) 
        if (y < maxy):
            maxy = y
            
    # Now we check the ORIGINAL LIST OF BOUNDING BOXES
    # to see if it matches the Y value
    starty = 0
    for cnt in contours:
        x, y, w, h = cv2.boundingRect(cnt) 
        if y >= maxy:
            # Drawing a rectangle on copied image 
            starty = y
            rect = cv2.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2) 


    mask = np.zeros(img.shape[:2], dtype="uint8")
    cv2.rectangle(mask, (0,starty), (width,height), (255,255,255), -1)
    masked = cv2.bitwise_and(newimg, newimg, mask=mask)

    # If necessary, further processing can be performed at this stage.

    cv2.imwrite(roiname,masked)
    cv2.imwrite(out,img)