Comparing Two Images in OpenCV Python: A Developer’s Guide

by

💡 Problem Formulation: You have two images and you need to determine how similar they are. Maybe you’re working on an automated system to detect duplicate images or verify if a photo matches a template. The input would be the two images in question and the desired output is a metric or method that quantitatively or qualitatively measures the similarity between these images.

Method 1: Mean Squared Error (MSE)

The Mean Squared Error (MSE) is a method that calculates the average squared difference between the pixel intensities of two images. It is used as a risk metric that quantifies the quality of an estimator; in this case, how similar two images are. A lower MSE means higher similarity. In OpenCV with Python, the function cv2.absdiff() is often employed followed by squaring and averaging the result.

Here’s an example:

import cv2
import numpy as np

def mse(imageA, imageB):
    err = np.sum((imageA.astype("float") - imageB.astype("float")) ** 2)
    err /= float(imageA.shape[0] * imageA.shape[1])
    
    return err

# Read two images
image1 = cv2.imread('image1.png')
image2 = cv2.imread('image2.png')

# Calculate MSE
mse_value = mse(image1, image2)

print(f"MSE: {mse_value}")

Output:

MSE: 350.4

The code defines a function mse() that takes two images, converts them to floating-point precision, subtracts them and squares the result, then sums all squared differences and finally normalizes by the number of pixels. We then read the two images, apply our MSE function, and the output is the Mean Squared Error between the images.

Method 2: Structural Similarity Index (SSI)

The Structural Similarity Index (SSI) is a perceptual metric that quantifies image quality degradation caused by processing. It is based on the computation of three components that constitute similarity: luminance, contrast, and structure. A value of 1 indicates perfect similarity. The SSI can be computed using OpenCV and the scikit-image library’s compare_ssim() function.

Here’s an example:

from skimage.metrics import structural_similarity as ssim
import cv2

# Read two images
image1 = cv2.imread('image1.png', cv2.IMREAD_GRAYSCALE)
image2 = cv2.imread('image2.png', cv2.IMREAD_GRAYSCALE)

# Calculate SSIM
ssi_value = ssim(image1, image2)

print(f"SSI: {ssi_value}")

Output:

SSI: 0.89

This snippet first loads in the two images in grayscale. It then calls the compare_ssim() function from scikit-image’s metrics module to compute the Structural Similarity Index. Finally, it prints the SSI value, representing the similarity between the two images.

Method 3: Feature Matching

Feature Matching involves detecting key points and their descriptors in both images, and then finding the best matches. This is particularly useful in object recognition and can accommodate differences in scale, rotation, and illumination. OpenCV provides several feature detectors and descriptor matchers, for example, the ORB (Oriented FAST and Rotated BRIEF) is a popular choice.

Here’s an example:

import cv2

# Initialize ORB detector
orb = cv2.ORB_create()

# Read two images
image1 = cv2.imread('image1.png')
image2 = cv2.imread('image2.png')

# Find keypoints and descriptors
kp1, des1 = orb.detectAndCompute(image1, None)
kp2, des2 = orb.detectAndCompute(image2, None)

# Create BFMatcher and find the matches
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
matches = bf.match(des1, des2)

# Sort matches by distance
matches = sorted(matches, key=lambda x: x.distance)

# Draw first ten matches
result = cv2.drawMatches(image1, kp1, image2, kp2, matches[:10], None, flags=2)

# Display the result
cv2.imshow('Feature Matching', result)
cv2.waitKey(0)
cv2.destroyAllWindows()

The snippet initializes an ORB detector, computes keypoints and descriptors, creates a matcher, finds the matches, sorts them, and then draws the top matches. The resulting image shows the connections between matched features, giving a visual representation of similarity.

Method 4: Histogram Comparison

Histogram Comparison is a method where histograms of both images are compared using various methods such as correlation, chi-square, intersection, or the Bhattacharyya distance. These histograms represent the distribution of pixel intensities. OpenCV provides the compareHist() function to facilitate this comparison.

Here’s an example:

import cv2

# Read images
image1 = cv2.imread('image1.png')
image2 = cv2.imread('image2.png')

# Convert to HSV color space
hsv1 = cv2.cvtColor(image1, cv2.COLOR_BGR2HSV)
hsv2 = cv2.cvtColor(image2, cv2.COLOR_BGR2HSV)

# Calculate Histograms
hist1 = cv2.calcHist([hsv1], [0, 1], None, [180, 256], [0, 180, 0, 256])
hist2 = cv2.calcHist([hsv2], [0, 1], None, [180, 256], [0, 180, 0, 256])

# Normalize the histograms
cv2.normalize(hist1, hist1, alpha=0, beta=1, norm_type=cv2.NORM_MINMAX)
cv2.normalize(hist2, hist2, alpha=0, beta=1, norm_type=cv2.NORM_MINMAX)

# Compare histograms using the correlation method
comparison = cv2.compareHist(hist1, hist2, cv2.HISTCMP_CORREL)

print(f"Histogram comparison: {comparison}")

Output:

Histogram comparison: 0.93

This code snippet reads two images, converts them to the HSV color space, and then computes their histograms. After normalization, it compares the histograms using the correlation method. The resulting value indicates how similar the color distribution is between the images.

Bonus One-Liner Method 5: Pixel-wise Absolute Difference

This is a simple approach where you calculate the absolute difference between two images on a pixel-by-pixel basis. This method quickly reveals areas with differing pixel values. In OpenCV, this can be achieved using the cv2.absdiff() function.

Here’s an example:

import cv2

# Read two images
image1 = cv2.imread('image1.png')
image2 = cv2.imread('image2.png')

# Calculate and show absolute difference
diff = cv2.absdiff(image1, image2)
cv2.imshow('Absolute Difference', diff)
cv2.waitKey(0)
cv2.destroyAllWindows()

The code loads two images and computes the absolute difference between them. The differing areas are made visible when displaying the diff image.

Summary/Discussion

  • Method 1: Mean Squared Error (MSE). Good for precise pixel-level comparison. Sensitive to noise and intensity variations.
  • Method 2: Structural Similarity Index (SSI). Mimics the human visual perception system. Not suitable for high detail comparison.
  • Method 3: Feature Matching. Excellent for object recognition and robust to transformations. May fail with textureless images or under extreme changes.
  • Method 4: Histogram Comparison. Great for global intensity distribution comparison. Not good for spatial comparison.
  • Bonus One-Liner Method 5: Pixel-wise Absolute Difference. Quick analysis of differences. Simple, but not robust to minor variations.