detecting articles, titles, and images in newspaper documents opencv

Question

I have a collection of newspapers were my goal is to get bounding boxes around several ROIs. Specifically these ROIs are:

• articles
• titles
• images

I am using opencv with python for achieving the desired result. My approach is as follows:

1. Canny Edge detection (proceeded by image blur)
2. Dilation
3. Contour Detection
4. Contour Approximation and Bound Box

I have written the appropriate code to achieve this, but the contours are not being so accurate (more on this will follow). Example: I am showing how the code is behaving:

As you can see, it is not detecting all the articles there, also it is combining parts of several articles into one bounding box. How can I make this better ? I want it to be more accurate. I played around with the parameters for Canny, Dilation, and blurring but did not achieve better results. Here is my code:

import cv2
import imutils
import numpy as np
import random

# capture image
path = 'C:/Users/96171/Desktop/dataset_training/jpg/75120201.jpg'
image = cv2.imread(path)
image = imutils.resize(image, width=500)
resize_factor = 1


blur = cv2.GaussianBlur(image, (9, 9), 0)
cv2.imshow('Blurred', blur)
cv2.waitKey()

# this was the golden line that made it better
# edged = cv2.Canny(blur, 0, 150)
edged = cv2.Canny(blur, 0, 170)

cv2.imshow('Edged', edged)
cv2.waitKey()

# edged = cv2.Canny(image, 0, 150)
# cv2.imshow("Edged", edged)
# cv2.waitKey()

# dilated = cv2.dilate(edged, np.ones((15, 15)))
dilated = cv2.dilate(edged, np.ones((3, 3)), iterations=1)
cv2.imshow('Dilated', dilated)
cv2.waitKey()


def _contour_approx_bad(contour, *args, **kwargs):
    """
    Approximate contour and discard non rectangular contours
    :returns: True if rectangle else False
    """
    perimeter = cv2.arcLength(contour, True)
    approx = cv2.approxPolyDP(contour, 0.02 * perimeter, True)

    return len(approx) == 4


contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
for i, contour in enumerate(contours):
    if not _contour_approx_bad(contour):
        rect = cv2.boundingRect(contour)
        x, y, w, h = [r*resize_factor for r in rect]
        b, g = random.sample(range(0, 255), 2)
        cv2.rectangle(image, (x,y), ((x+w), (y+h)), (b, g, 255), 3)
        # self.crop(name=str(i), **{'start': (x,y), 'end': ((x+w), (y+h))})

cv2.imshow('Final Image', image)
cv2.waitKey()
cv2.imwrite('tobecropped.png', image)