Spaces:
Sleeping
Sleeping
| from PIL import Image, ImageDraw, ImageFont | |
| import numpy as np | |
| import cv2 | |
| import os | |
| FONTS_FOLDER = "fonts" | |
| NUM_IMAGES_PER_FONT = 5 | |
| def generate_images(text): | |
| images = [] | |
| for font_file in os.listdir(FONTS_FOLDER): | |
| font_path = os.path.join(FONTS_FOLDER, font_file) | |
| for i in range(NUM_IMAGES_PER_FONT): | |
| img = generate_text_image(text, font_path) | |
| images.append((img, font_file)) | |
| return images | |
| def generate_text_image(text, font_path, fontsize=None): | |
| if not fontsize: | |
| fontsize = int(np.random.normal(loc=50, scale=10)) | |
| font = ImageFont.truetype(font_path, fontsize) | |
| text_size = font.getsize(text) | |
| img = Image.new('RGB', text_size, color='black') | |
| draw = ImageDraw.Draw(img) | |
| draw.text((0, 0), text, font=font, fill='white') | |
| noise = np.random.normal(loc=0, scale=10, size=(img.size[1], img.size[0]))[..., np.newaxis] | |
| noise = np.tile(noise, [1, 1, 3]) | |
| img = Image.fromarray(np.clip(np.array(img) + noise, 0, 255).astype(np.uint8), 'RGB') | |
| return np.array(img) | |
| def flann_matching_alt(generated_images, query_image, num_trees=5, num_checks=50): | |
| query_image_gray = cv2.cvtColor(query_image, cv2.COLOR_BGR2GRAY) | |
| generated_images_gray = [] | |
| for img, _ in generated_images: | |
| generated_images_gray.append(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)) | |
| sift = cv2.SIFT_create() | |
| index_params = dict(algorithm=0, trees=num_trees) | |
| search_params = dict(checks=num_checks) | |
| flann = cv2.FlannBasedMatcher(index_params, search_params) | |
| query_kp, query_desc = sift.detectAndCompute(query_image_gray, None) | |
| if query_desc is None: | |
| return [] | |
| matching_results = [] | |
| for i, (img, font_file) in enumerate(generated_images): | |
| kp, desc = sift.detectAndCompute(generated_images_gray[i], None) | |
| if desc is not None: | |
| matches = flann.knnMatch(query_desc, desc, k=2) | |
| good_matches = [] | |
| for m, n in matches: | |
| if m.distance < 0.75 * n.distance: | |
| good_matches.append(m) | |
| matching_img = cv2.drawMatches(query_image_gray, query_kp, generated_images_gray[i], kp, good_matches, None, flags=2) | |
| # Calculate percentage match | |
| num_query_kp = len(query_kp) | |
| num_matches = len(good_matches) | |
| match_percent = 100 * num_matches / num_query_kp | |
| matching_results.append((matching_img, font_file, match_percent)) | |
| return matching_results | |
| def flann_matching(generated_images, query_image, num_trees=5, num_checks=50): | |
| query_image_gray = cv2.cvtColor(query_image, cv2.COLOR_BGR2GRAY) | |
| generated_images_gray = [] | |
| for img, _ in generated_images: | |
| generated_images_gray.append(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)) | |
| sift = cv2.SIFT_create() | |
| index_params = dict(algorithm=0, trees=num_trees) | |
| search_params = dict(checks=num_checks) | |
| flann = cv2.FlannBasedMatcher(index_params, search_params) | |
| query_kp, query_desc = sift.detectAndCompute(query_image_gray, None) | |
| generated_kp = [] | |
| generated_desc = [] | |
| for img in generated_images_gray: | |
| kp, desc = sift.detectAndCompute(img, None) | |
| generated_kp.append(kp) | |
| generated_desc.append(desc) | |
| matching_results = [] | |
| for i, (img, font_file) in enumerate(generated_images): | |
| matches = flann.knnMatch(query_desc, generated_desc[i], k=2) | |
| good_matches = [] | |
| for m, n in matches: | |
| if m.distance < 0.75*n.distance: | |
| good_matches.append([m]) | |
| matching_img = cv2.drawMatchesKnn(query_image_gray, query_kp, img, generated_kp[i], good_matches, None, flags=2) | |
| # Calculate percentage match | |
| num_query_kp = len(query_kp) | |
| num_matches = len(good_matches) | |
| match_percent = 100*num_matches/num_query_kp | |
| matching_results.append((matching_img, font_file, match_percent)) | |
| return matching_results | |
| def flann_matching_3(generated_images, query_image, num_trees=5, num_checks=50): | |
| query_image_gray = cv2.cvtColor(query_image, cv2.COLOR_BGR2GRAY) | |
| generated_images_gray = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img, _ in generated_images] | |
| sift = cv2.SIFT_create() | |
| index_params = dict(algorithm=0, trees=num_trees) | |
| search_params = dict(checks=num_checks) | |
| flann = cv2.FlannBasedMatcher(index_params, search_params) | |
| query_kp, query_desc = sift.detectAndCompute(query_image_gray, None) | |
| if query_desc is None: | |
| return [] | |
| matching_results = [] | |
| for i, (img, font_file) in enumerate(generated_images): | |
| kp, desc = sift.detectAndCompute(generated_images_gray[i], None) | |
| if desc is None: | |
| continue | |
| matches = flann.knnMatch(query_desc, desc, k=2) | |
| good_matches = [] | |
| for m, n in matches: | |
| if m.distance < 0.75 * n.distance: | |
| good_matches.append(m) | |
| if len(good_matches) < 10: | |
| continue | |
| src_pts = np.float32([query_kp[m.queryIdx].pt for m in good_matches]).reshape(-1, 1, 2) | |
| dst_pts = np.float32([kp[m.trainIdx].pt for m in good_matches]).reshape(-1, 1, 2) | |
| M, _ = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0) | |
| if M is None: | |
| continue | |
| h, w = query_image_gray.shape | |
| dst_img = cv2.warpPerspective(img, M, (w, h)) | |
| dst_gray = cv2.cvtColor(dst_img, cv2.COLOR_BGR2GRAY) | |
| orb = cv2.ORB_create() | |
| kp1, desc1 = sift.detectAndCompute(query_image_gray, None) | |
| kp2, desc2 = sift.detectAndCompute(dst_gray, None) | |
| if desc1 is None or desc2 is None: | |
| continue | |
| matches = flann.knnMatch(desc1, desc2, k=2) | |
| good_matches = [] | |
| for m, n in matches: | |
| if m.distance < 0.75 * n.distance: | |
| good_matches.append(m) | |
| if len(good_matches) < 10: | |
| continue | |
| matching_img = cv2.drawMatches(query_image_gray, kp1, dst_gray, kp2, good_matches, None, flags=2) | |
| # Calculate percentage match | |
| num_query_kp = len(kp1) | |
| num_matches = len(good_matches) | |
| match_percent = 100 * num_matches / num_query_kp | |
| matching_results.append((matching_img, font_file, match_percent)) | |
| return matching_results | |
| def flann_matching_4(generated_images, query_image, num_trees=5, num_checks=50): | |
| query_image_gray = cv2.cvtColor(query_image, cv2.COLOR_BGR2GRAY) | |
| generated_images_gray = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img, _ in generated_images] | |
| sift = cv2.SIFT_create() | |
| index_params = dict(algorithm=0, trees=num_trees) | |
| search_params = dict(checks=num_checks) | |
| flann = cv2.FlannBasedMatcher(index_params, search_params) | |
| query_kp, query_desc = sift.detectAndCompute(query_image_gray, None) | |
| if query_desc is None: | |
| return [] | |
| matching_results = [] | |
| for img, font_file in generated_images: | |
| generated_image_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) | |
| kp, desc = sift.detectAndCompute(generated_image_gray, None) | |
| if desc is None: | |
| continue | |
| matches = flann.knnMatch(query_desc, desc, k=2) | |
| good_matches = [] | |
| for m, n in matches: | |
| if m.distance < 0.75 * n.distance: | |
| good_matches.append(m) | |
| matching_img = cv2.drawMatches(query_image_gray, query_kp, generated_image_gray, kp, good_matches, None, flags=2) | |
| # Calculate percentage match | |
| num_query_kp = len(query_kp) | |
| num_matches = len(good_matches) | |
| match_percent = 100 * num_matches / num_query_kp | |
| matching_results.append((matching_img, font_file, match_percent)) | |
| return matching_results |