sift.py

import matplotlib.pyplot as plt
import cv2
import os

# Ensures no GUI required for Qt, may cause errors
os.environ['QT_QPA_PLATFORM'] = 'offscreen'

def extract_sift_features(img):
	""" Extracts SIFT features from the given image """
	# SIFT algorithm is stored in the sift_initialize variable
	sift_initialize = cv2.SIFT_create()
	key_points, descriptors = sift_initialize.detectAndCompute(img, None)
	return key_points, descriptors

def showing_sift_features(img1, img2, key_points):
	"""
	Draws the keypoints present in the image
	:param key_points: key points between :param img1: and :param img2:
	"""
	# Convert to RGB for display
	img1_rgb = cv2.cvtColor(img1, cv2.COLOR_BGR2RGB)
	plt.imshow(cv2.drawKeypoints(img1_rgb, key_points, img2.copy()))

def sift(self, filename1, filename2, out_filename, nr_matches=100):
	"""
	SIFT algorithm
	:param filename1: filename of the first image
	:param filename2: filename of the second image 
	:param out_filename: output filename
	The algorithm will show the first :param nr_matches:
	"""
	image1 = cv2.imread(filename1)
	image2 = cv2.imread(filename2)

	h1, w1 = image1.shape[:2]
	h2, w2 = image2.shape[:2]

	if h1 * w1 > h2 * w2:
		image1 = cv2.resize(image1, (w2, h2))
	else:
		image2 = cv2.resize(image2, (w1, h1)) 

	# Convert images to grayscale for SIFT
	image1_gray = cv2.cvtColor(image1, cv2.COLOR_BGR2GRAY)
	image2_gray = cv2.cvtColor(image2, cv2.COLOR_BGR2GRAY)

	# Extract key points and descriptors
	image1_key_points, image1_descriptors = extract_sift_features(image1_gray)
	image2_key_points, image2_descriptors = extract_sift_features(image2_gray)

	showing_sift_features(image1_gray, image1, image1_key_points)

	""" Match descriptors using the norm. This could be problematic because
		it finds the most common pixels, even if the images are completely
		different (could be the corners/borders that are similar).
		Find distance between key points using the Manhattan distance (i.e. norm)
	"""
	bruteForce = cv2.BFMatcher(cv2.NORM_L2)

	# Matches descriptors
	matches = bruteForce.match(image1_descriptors, image2_descriptors)

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

	# Save only the first 100 matches
	matched_img = cv2.drawMatches(image1, image1_key_points, 
								image2, image2_key_points, 
								matches[:nr_matches], image2.copy())
	matched_img_rgb = cv2.cvtColor(matched_img, cv2.COLOR_BGR2RGB)

	plt.figure(figsize=(30, 15))
	plt.imshow(matched_img_rgb)
	plt.axis('off')
	plt.savefig(out_filename)

def attach_sift_to_image(image_class):
	"""
	Attaches the method to the given image class.
	:param image_class: The class to which the method will be attached.
	"""
	image_class.sift = sift