Function to detect fields from a form, may need fine tuning!

contour_field_detection.py

from PIL import Image
import numpy as np
import cv2
import time

# Helper function for jupyter notebooks
def showarray(a):
    return Image.fromarray(a)

def detect_boxes(img, min_percent_of_area=0.02, max_percent_of_area=2):
    
    # Make a copy of the image 
    cp_img = img.copy()
    gray = cv2.cvtColor(cp_img, cv2.COLOR_RGB2GRAY)
    h, w, _ = img.shape
    blank_image = np.zeros((h,w), np.uint8)

    # Use edge detection to pull out lines, notice aperture size is not default, 5 works best
    edges = cv2.Canny(gray,0, 10, apertureSize=5)

    # Detect contours using the default settings
    _, contours,_ = cv2.findContours(edges,1,2)

    # If we find any contours that have 4 corners and within restrictions then add them onto the blank image
    for cnt in contours:
        approx = cv2.approxPolyDP(cnt,0.02*cv2.arcLength(cnt,True),True)
        if len(approx) == 4:
            if cv2.contourArea(cnt) > min_percent_of_area*h*w/100 and cv2.contourArea(cnt) < max_percent_of_area*h*w/100:
                cv2.drawContours(blank_image,[cnt],0,(255),-1)

    # Find all of the contours on the new image we have formed, this time with no restrictions, this now gets
    #   one box if the contours were overlapping
    _, contours,_ = cv2.findContours(blank_image,1,2)
    
    # Now get the miniumum bounding box for each of the contours and return them as top_left and bottom_right
    countours_bb = [cv2.boundingRect(cnt) for cnt in contours]
    rects = [[(cnt[0], cnt[1]), (cnt[0]+cnt[2], cnt[1]+cnt[3])] for cnt in countours_bb]
    
    return rects
  
if __name__ == '__main__':
    
    IMAGE_PATH = 'path/to/image.png'
    
    # Read image in
    img = cv2.imread(IMAGE_PATH)
    
    # Detect fields
    rects = detect_boxes(img)
    
    # Draw predicted boxes on
    _ = [cv2.rectangle(img, rect[0], rect[1], (255,0,0), 2) for rect in rects]
    
    # Show image in Jupyter Notebook
    showarray(img)