Python | Document field detection using Template Matching

Template matching is an image processing technique which is used to find the location of small-parts/template of a large image. This technique is widely used for object detection projects, like product quality, vehicle tracking, robotics etc.
In this article, we will learn how to use template matching for detecting the related fields in a document image.
Solution:
Above task can be achieved using template matching. Clip out the field images and apply template matching using clipped field images and the document image. The algorithm is simple yet reproducible into complex versions to solve the problem of field detection and localization for document images belonging to specific domains.
Approach:

Clip/Crop field images from the main document and use them as separate templates.
Define/tune thresholds for different fields.
Apply template matching for each cropped field template using OpenCV function cv2.matchTemplate()
Draw bounding boxes using the coordinates of rectangles fetched from template matching.
Optional: Augment field templates and fine tune threshold to improve result for different document images.

Input Image:

Output Image:

Below is the Python code:

Python3

# importing libraries 

import numpy as np 

import imutils 

import cv2 

field_threshold = { "prev_policy_no" : 0.7, 

                    "address"        : 0.6, 

                  } 

# Function to Generate bounding 
# boxes around detected fields 

def getBoxed(img, img_gray, template, field_name = "policy_no"): 

    w, h = template.shape[::-1]  

    # Apply template matching 

    res = cv2.matchTemplate(img_gray, template, 

                           cv2.TM_CCOEFF_NORMED) 

    hits = np.where(res >= field_threshold[field_name]) 

    # Draw a rectangle around the matched region.  

    for pt in zip(*hits[::-1]):  

        cv2.rectangle(img, pt, (pt[0] + w, pt[1] + h), 

                                    (0, 255, 255), 2) 

        y = pt[1] - 10 if pt[1] - 10 > 10 else pt[1] + h + 20

        cv2.putText(img, field_name, (pt[0], y), 

            cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 1) 

    return img 

# Driver Function 

if __name__ == '__main__': 

    # Read the original document image 

    img = cv2.imread('doc.png') 

    # 3-d to 2-d conversion 

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

    # Field templates 

    template_add = cv2.imread('doc_address.png', 0) 

    template_prev = cv2.imread('doc_prev_policy.png', 0) 

    img = getBoxed(img.copy(), img_gray.copy(), 

                       template_add, 'address') 

    img = getBoxed(img.copy(), img_gray.copy(), 

                   template_prev, 'prev_policy_no') 

    cv2.imshow('Detected', img)