Multi-Threaded SIFT feature comparison
* Comparison based on biggest SIFT features * Multi-threaded comparison of one image against a set * Basic folder structure for importing new images into the setmaster
parent
276db4b9fb
commit
32fcc04c14
@ -0,0 +1,178 @@
|
||||
import os
|
||||
import cv2
|
||||
from ImageComparator import ImageComparator
|
||||
from datetime import datetime
|
||||
|
||||
|
||||
class ImageDB:
|
||||
|
||||
supported_file_extensions = [".png", ".jpg", ".jpeg"]
|
||||
|
||||
def __init__(self, root_path: str, import_folder: str = "import", db_folder: str = "images",
|
||||
dump_folder: str = "duplicate",
|
||||
folder_sep: str = "/", samples: int = 100, dist_thresh: float = 80,
|
||||
match_thresh: float = 0.3, resize_dim: tuple = (500,500), threads: int = 2):
|
||||
self.start_time = datetime.now()
|
||||
|
||||
self.root_path = root_path
|
||||
self.import_path = self.root_path + folder_sep + import_folder
|
||||
self.db_path = self.root_path + folder_sep + db_folder
|
||||
self.dump_path = self.root_path + folder_sep + dump_folder
|
||||
self.dir_sep = folder_sep
|
||||
|
||||
log_file = self.root_path + self.dir_sep + "db_log.txt"
|
||||
|
||||
self.imgcomp = ImageComparator(samples, dist_thresh, match_thresh, resize_dim, threads)
|
||||
|
||||
if not os.path.isdir(self.root_path):
|
||||
os.mkdir(self.root_path)
|
||||
self.log("Created {0} because it was missing ...".format(self.root_path), False)
|
||||
if not os.path.isdir(self.import_path):
|
||||
os.mkdir(self.import_path)
|
||||
self.log("Created {0} because it was missing ...".format(self.import_path), False)
|
||||
if not os.path.isdir(self.db_path):
|
||||
os.mkdir(self.db_path)
|
||||
self.log("Created {0} because it was missing ...".format(self.db_path), False)
|
||||
if not os.path.isdir(self.dump_path):
|
||||
os.mkdir(self.dump_path)
|
||||
self.log("Created {0} because it was missing ...".format(self.dump_path), False)
|
||||
|
||||
self.log_stream = open(log_file, "w")
|
||||
|
||||
self.log("Calculating all features ...")
|
||||
self.picture_data = self.calc_db_features()
|
||||
self.log("Done!")
|
||||
|
||||
def log(self, msg: str, log_to_file: bool = True):
|
||||
time = datetime.now() - self.start_time
|
||||
msg = "[{0}] {1}".format(time, msg)
|
||||
print(msg)
|
||||
if log_to_file:
|
||||
self.log_stream.write(msg+"\n")
|
||||
self.log_stream.flush()
|
||||
|
||||
def process_all_images(self):
|
||||
for file_name in os.listdir(self.import_path):
|
||||
self.process_image(file_name)
|
||||
self.update_db_features()
|
||||
|
||||
def process_image(self, img_name: str) -> bool:
|
||||
"""
|
||||
Processes a given image and decides if it is to be imported and then import or return.
|
||||
|
||||
Assumes image to be inside the import folder.
|
||||
Indicates if it was imported via boolean
|
||||
:param img_name:
|
||||
:return:
|
||||
"""
|
||||
img_path = self.import_path + self.dir_sep + img_name
|
||||
|
||||
if self.is_supported_file(img_path):
|
||||
result, matched, ratio = self.is_image_unique(img_path, self.picture_data)
|
||||
if not result:
|
||||
self.import_image(img_name)
|
||||
return True
|
||||
else:
|
||||
self.log("Matched {0} to {1} with ratio {2}".format(img_name, matched, ratio))
|
||||
self.move_to_dump(img_name)
|
||||
return False
|
||||
|
||||
def import_image(self, img_name: str):
|
||||
"""
|
||||
Import the given image into the database.
|
||||
|
||||
Does no checks, just imports.
|
||||
:param img_name: image file name, assumed to be located inside the import folder
|
||||
:return:
|
||||
"""
|
||||
img_path = self.import_path + self.dir_sep + img_name
|
||||
dest_path = self.db_path + self.dir_sep + img_name
|
||||
if os.path.isfile(img_path):
|
||||
os.rename(img_path, dest_path)
|
||||
else:
|
||||
raise Exception("Image path has not been a valid file!")
|
||||
|
||||
def move_to_dump(self, img_name: str):
|
||||
"""
|
||||
Move the given image to the dump folder.
|
||||
:param img_name:
|
||||
:return:
|
||||
"""
|
||||
img_path = self.import_path + self.dir_sep + img_name
|
||||
dest_path = self.dump_path + self.dir_sep + img_name
|
||||
if os.path.isfile(img_path):
|
||||
os.rename(img_path, dest_path)
|
||||
else:
|
||||
raise Exception("Image path has not been a valid file!")
|
||||
|
||||
def is_image_unique(self, img_path: str, picture_data: dict) -> tuple:
|
||||
"""
|
||||
Check if the given image is already in the databank or not.
|
||||
:param img_path:
|
||||
:param picture_data:
|
||||
:return:
|
||||
"""
|
||||
if len(picture_data) == 0:
|
||||
return False, "", 0.0
|
||||
result, matched, match = self.imgcomp.has_similar_match(img_path, self.db_path,
|
||||
picture_data)
|
||||
|
||||
if result:
|
||||
return True, matched, match
|
||||
else:
|
||||
return False, "", 0.0
|
||||
|
||||
def calc_db_features(self) -> dict:
|
||||
"""
|
||||
Calculate keypoints for every image in the database
|
||||
:return:
|
||||
"""
|
||||
pictures = dict()
|
||||
for name in os.listdir(self.db_path):
|
||||
pic_path = self.db_path + self.dir_sep + name
|
||||
if not self.is_supported_file(pic_path):
|
||||
continue
|
||||
img = cv2.imread(pic_path)
|
||||
|
||||
pictures[pic_path] = self.imgcomp.get_features(img)
|
||||
|
||||
return pictures
|
||||
|
||||
def update_db_features(self):
|
||||
"""
|
||||
Update the keypoint dictionary with only new image data
|
||||
:return:
|
||||
"""
|
||||
for name in os.listdir(self.db_path):
|
||||
pic_path = self.db_path + self.dir_sep + name
|
||||
if not self.is_supported_file(pic_path):
|
||||
continue
|
||||
if pic_path not in self.picture_data.keys():
|
||||
self.log("New image {0}! Calculating features for memory...".format(name))
|
||||
img = cv2.imread(pic_path)
|
||||
self.picture_data[pic_path] = self.imgcomp.get_features(img)
|
||||
self.log("Done!")
|
||||
|
||||
def is_supported_file(self, path: str) -> bool:
|
||||
"""
|
||||
Returns if the given file has a valid picture file extension.
|
||||
:param path:
|
||||
:return:
|
||||
"""
|
||||
if os.path.isfile(path):
|
||||
for ext in ImageDB.supported_file_extensions:
|
||||
if ext in path.lower():
|
||||
return True
|
||||
print("{0} is not a supported file format!".format(path))
|
||||
return False
|
||||
|
||||
def get_db_size(self) -> int:
|
||||
"""
|
||||
Return the number of images inside the databank
|
||||
:return:
|
||||
"""
|
||||
size = 0
|
||||
for entry in os.listdir(self.db_path):
|
||||
if os.path.isfile(self.db_path+self.dir_sep+entry):
|
||||
size += 1
|
||||
return size
|
||||
@ -0,0 +1,168 @@
|
||||
import cv2
|
||||
from ImageCompareThread import ImageCompareManageThread
|
||||
|
||||
import datetime
|
||||
|
||||
|
||||
class ImageComparator:
|
||||
|
||||
concurrent_threads = 10
|
||||
|
||||
def __init__(self, samples: int = 100, dist_thresh: float = 80, match_thresh: float = 0.6,
|
||||
resize_dim: tuple = (500, 500), concurrent_threads: int = 10):
|
||||
self.results = {}
|
||||
ImageComparator.concurrent_threads = concurrent_threads
|
||||
self.icmt = None
|
||||
self.samples = samples
|
||||
self.dist_thresh = dist_thresh
|
||||
self.match_thresh = match_thresh
|
||||
self.resize_dim = resize_dim
|
||||
|
||||
self.sift = cv2.xfeatures2d.SIFT_create()
|
||||
|
||||
@DeprecationWarning
|
||||
def match_images(self, path0: str, path1: str, sample_size: int = 100,
|
||||
match_thresh: float = 0.8,
|
||||
dist_thresh: float = 350,
|
||||
diff_min: float = 1.5) -> bool:
|
||||
"""
|
||||
Matches the given images using SIFT featuring and euclidian distance comparison
|
||||
of a random sample of keypoints.
|
||||
|
||||
True if at least match_thresh many keypoints
|
||||
have been successfully matched (e.g. 0.9 -> 90%).
|
||||
:param path0:
|
||||
:param path1:
|
||||
:param sample_size:
|
||||
:param match_thresh: float from 0 to 1 in percent of keypoint matches required
|
||||
:param dist_thresh: float in max distance of keypoints to match
|
||||
:param diff_min: float by which the second closest match must be bigger
|
||||
:return:
|
||||
"""
|
||||
start_time = datetime.datetime.now()
|
||||
print("Creating feature lists ...")
|
||||
ft0, des0 = self.get_features(cv2.imread(path0))
|
||||
ft1, des1 = self.get_features(cv2.imread(path1))
|
||||
print("Created feature lists!")
|
||||
|
||||
print(datetime.datetime.now() - start_time)
|
||||
|
||||
print("Looking for matches ...")
|
||||
selection = self.__get_random_selection(des0, sample_size)
|
||||
hits = self.find_matching_keypoints(des0, des1, sample_size, dist_thresh)
|
||||
print("Looked for matches!")
|
||||
|
||||
print(datetime.datetime.now() - start_time)
|
||||
|
||||
match_ratio = hits / len(selection)
|
||||
print("MatchRatio:{0} Hits:{1}".format(match_ratio, hits))
|
||||
|
||||
if match_ratio >= match_thresh:
|
||||
return True
|
||||
else:
|
||||
return False
|
||||
|
||||
@DeprecationWarning
|
||||
def find_matching_keypoints(self, keypoints1: list, keypoints2: list,
|
||||
sample_size: int, dist_thresh: float) -> int:
|
||||
"""
|
||||
Find nearest neighbours for each point in keypoints1 in keypoints2.
|
||||
|
||||
Returns number of sufficiently matching keypoints
|
||||
:param keypoints1:
|
||||
:param keypoints2:
|
||||
:param sample_size:
|
||||
:param dist_thresh:
|
||||
:return:
|
||||
"""
|
||||
selection = self.__get_random_selection(keypoints1, sample_size)
|
||||
hits, sum = 0, 0
|
||||
for i in range(len(selection)):
|
||||
hit, dist = self.has_matching_keypoint(selection[i], keypoints2, dist_thresh)
|
||||
if hit:
|
||||
hits += 1
|
||||
sum += dist
|
||||
|
||||
return hits
|
||||
|
||||
def get_features(self, img) -> tuple:
|
||||
assert(img is not None)
|
||||
grey = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
|
||||
|
||||
kp = self.sift.detect(grey, None)
|
||||
kp = self.select_keypoints(kp, self.samples, (img.shape[1], img.shape[0]))
|
||||
des = self.sift.compute(grey, kp)
|
||||
|
||||
return kp, des[1]
|
||||
|
||||
def select_keypoints(self, kps: list, samples: int, dimension: tuple):
|
||||
"""
|
||||
Select a good sample of keypoints among all keypoints.
|
||||
:param kps:
|
||||
:return:
|
||||
"""
|
||||
result = []
|
||||
kps.sort(key=lambda x: x.pt[0]) # sort by x-coord
|
||||
|
||||
kps_left = kps[:len(kps)//2]
|
||||
kps_left.sort(key=lambda x: x.pt[1])
|
||||
|
||||
kps_up_left = kps_left[:len(kps_left)//2]
|
||||
kps_down_left = kps_left[len(kps_left)//2:]
|
||||
|
||||
kps_right = kps[len(kps)//2:]
|
||||
kps_right.sort(key=lambda x: x.pt[1])
|
||||
|
||||
kps_up_right = kps_right[:len(kps_right)//2]
|
||||
kps_down_right = kps_right[len(kps_right)//2:]
|
||||
|
||||
sample = samples // 4
|
||||
|
||||
for quad_kps in [kps_up_left, kps_up_right, kps_down_left, kps_down_right]:
|
||||
quad_kps.sort(key=lambda x: x.response, reverse=True)
|
||||
|
||||
result += quad_kps[:sample]
|
||||
|
||||
return result
|
||||
|
||||
def has_similar_match(self, imgPath: str, dbPath: str, picture_data: dict) -> tuple:
|
||||
"""
|
||||
Calculate similarity of imgPath and all images in dbPath.
|
||||
Returns as soon as a match has been found.
|
||||
:param imgPath:
|
||||
:param dbPath:
|
||||
:param picture_data:
|
||||
:return:
|
||||
"""
|
||||
# Calculating features
|
||||
targImg = cv2.imread(imgPath)
|
||||
targFeat, targDesc = self.get_features(targImg)
|
||||
|
||||
# Matching features
|
||||
|
||||
self.icmt = ImageCompareManageThread(imgPath, targDesc, picture_data, self.notify_result,
|
||||
self.concurrent_threads, self.samples,
|
||||
self.dist_thresh)
|
||||
self.icmt.start()
|
||||
self.icmt.join()
|
||||
print("Done managing threads")
|
||||
|
||||
match = ""
|
||||
value = 0.0
|
||||
for key in self.results.keys():
|
||||
if value < self.results[key]/self.samples:
|
||||
value = self.results[key] / self.samples
|
||||
match = key
|
||||
|
||||
if value >= self.match_thresh:
|
||||
return True, match, value
|
||||
else:
|
||||
return False, "", 0.0
|
||||
|
||||
def notify_result(self, name: str, hits: int):
|
||||
self.results[name] = hits
|
||||
|
||||
if hits/self.samples >= self.match_thresh:
|
||||
print("[{0}] got a match! Aborting search ...".format(name))
|
||||
self.icmt.searching = False
|
||||
|
||||
@ -0,0 +1,121 @@
|
||||
from threading import Thread
|
||||
from time import sleep
|
||||
import numpy as np
|
||||
import random
|
||||
from scipy.spatial import KDTree
|
||||
|
||||
|
||||
class ImageCompareManageThread(Thread):
|
||||
|
||||
def __init__(self, name: str, candidate_desc: np.array, descriptors: dict, callback,
|
||||
concurrent_threads: int = 10, samples: int = 100, dist_thresh: float = 80):
|
||||
super().__init__(name=name)
|
||||
|
||||
self.candidate_desc = candidate_desc
|
||||
self.descriptors = descriptors
|
||||
self.samples = samples
|
||||
self.dist_thresh = dist_thresh
|
||||
self.callback = callback
|
||||
self.concurrent_threads = concurrent_threads
|
||||
|
||||
self.todo = []
|
||||
self.threads = {}
|
||||
self.searching = False
|
||||
|
||||
def run(self):
|
||||
print("[{0}] Starting management ...".format(self.name))
|
||||
self.searching = True
|
||||
self.todo = list(self.descriptors.keys())
|
||||
|
||||
for i in range(self.concurrent_threads):
|
||||
if len(self.todo) == 0:
|
||||
break
|
||||
key = self.todo.pop()
|
||||
ict = ImageCompareThread(key, self.candidate_desc, self.descriptors[key][1],
|
||||
self.samples, self.dist_thresh, self.finish_thread)
|
||||
self.threads[key] = ict
|
||||
ict.start()
|
||||
|
||||
while self.searching:
|
||||
sleep(2)
|
||||
|
||||
def finish_thread(self, name: str, hits: int):
|
||||
self.callback(name, hits)
|
||||
print("[{0}] finished with {1}".format(name, hits))
|
||||
|
||||
print("{0} jobs left ...".format(len(self.todo)))
|
||||
if len(self.todo) > 0 and self.searching: # still work to do, start another thread
|
||||
key = self.todo.pop()
|
||||
ict = ImageCompareThread(key, self.candidate_desc, self.descriptors[key][1],
|
||||
self.samples, self.dist_thresh, self.finish_thread)
|
||||
self.threads[key] = ict
|
||||
ict.start()
|
||||
else:
|
||||
self.searching = False
|
||||
|
||||
|
||||
class ImageCompareThread(Thread):
|
||||
|
||||
def __init__(self, name: str, candidate_desc: np.array, db_desc: np.array,
|
||||
sample_size: int, dist_thresh: float, callback):
|
||||
super().__init__(name=name)
|
||||
|
||||
self.candidate_desc = candidate_desc
|
||||
self.db_desc = db_desc
|
||||
self.samples = sample_size
|
||||
self.dist_thresh = dist_thresh
|
||||
self.callback = callback
|
||||
|
||||
def run(self):
|
||||
print("[{0}] starting ...".format(self.name))
|
||||
hits = self.find_matching_keypoints(self.candidate_desc, self.db_desc,
|
||||
self.samples, self.dist_thresh)
|
||||
|
||||
self.callback(self.name, hits)
|
||||
|
||||
def find_matching_keypoints(self, keypoints1: list, keypoints2: list,
|
||||
sample_size: int, dist_thresh: float) -> int:
|
||||
"""
|
||||
Find nearest neighbours for each point in keypoints1 in keypoints2.
|
||||
|
||||
Returns number of sufficiently matching keypoints
|
||||
:param keypoints1:
|
||||
:param keypoints2:
|
||||
:param sample_size:
|
||||
:param dist_thresh:
|
||||
:return:
|
||||
"""
|
||||
hits, sum = 0, 0
|
||||
for i in range(len(keypoints1)):
|
||||
hit, dist = self.has_matching_keypoint(keypoints1[i], keypoints2, dist_thresh)
|
||||
if hit:
|
||||
hits += 1
|
||||
sum += dist
|
||||
|
||||
return hits
|
||||
|
||||
def __get_random_selection(self, l: list, num: int) -> list:
|
||||
result = []
|
||||
for k in range(num):
|
||||
result.append(random.choice(l))
|
||||
|
||||
return result
|
||||
|
||||
def has_matching_keypoint(self, point: np.ndarray, points: list, max_dist: float) -> tuple:
|
||||
"""
|
||||
Find nearest neighbour for point in points.
|
||||
:param point:
|
||||
:param points:
|
||||
:param max_dist:
|
||||
:return:
|
||||
"""
|
||||
tree = KDTree(points)
|
||||
|
||||
dist, ind = tree.query([point], k=2)
|
||||
dist = dist[0] # just resolving nested lists
|
||||
#print("Distances:{0} Indexes:{1} MaxDist:{2}".format(dist, ind[0], max_dist))
|
||||
|
||||
if dist[0] <= max_dist: # second neighbour is found, valid hit
|
||||
return True, dist[0]
|
||||
else:
|
||||
return False, dist[0]
|
||||
@ -0,0 +1,22 @@
|
||||
from CompDatabase import ImageDB
|
||||
from datetime import datetime
|
||||
|
||||
start_time = datetime.now()
|
||||
# TODO save & load feature lists for the database, fix resized matching (injective resizing?)
|
||||
|
||||
root_folder = "Pictures"
|
||||
import_folder = "import"
|
||||
db_folder = "images"
|
||||
samples = 200
|
||||
dist_thresh = 80
|
||||
match_thresh = 0.8
|
||||
resize_dim = (500, 500)
|
||||
threads = 4
|
||||
|
||||
img_db = ImageDB(root_folder, import_folder, db_folder, samples=samples, dist_thresh=dist_thresh,
|
||||
match_thresh=match_thresh, resize_dim=resize_dim, threads=threads)
|
||||
|
||||
img_db.log("Starting work ...")
|
||||
img_db.log("DB size: {0}".format(img_db.get_db_size()))
|
||||
img_db.process_all_images()
|
||||
img_db.log("Work done!")
|
||||
Loading…
Reference in New Issue