Is there a way to run OpenCV's SIFT faster?
Question:
I have a directory of images that contains many unidentified duplicates. My goal is to identify the duplicates. Because the duplicates have been cropped, resized, or converted to a different image format, they cannot be detected by comparing their hashes.
I wrote a script that successfully detects duplicates, but there is one major drawback: the script is slow. On a test-drive with a folder containing 60 items, it took five hours to run (this might also be a reflection of my increasingly buggy and slow computer). Since I have approximately 66,000 images in my directory, I estimate that it will take 229 days for the script to complete.
Can anyone suggest solutions? My research has revealed that you can free up memory by "releasing" the image stored in the variable as the loop completes, but all the information on how to do this seems to be written in C, not python. I was also thinking of trying to use orb instead of sift, but have concerns about its accuracy. Does anyone have advice on which of the two options would be better to pursue? Or a way to rewrite the script so it take less memory? Many thanks in advance.
from __future__ import division
import cv2
import numpy as np
import glob
import pandas as pd
listOfTitles1 = []
listOfTitles2 = []
listOfSimilarities = []
# Sift and Flann
sift = cv2.xfeatures2d.SIFT_create()
index_params = dict(algorithm=0, trees=5)
search_params = dict()
flann = cv2.FlannBasedMatcher(index_params, search_params)
# Load all the images1
countInner = 0
countOuter = 1
folder = r"/Downloads/images/**/*"
for a in glob.iglob(folder,recursive=True):
for b in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if not b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countInner += 1
print(countInner, "", countOuter)
if countInner <= countOuter:
continue
image1 = cv2.imread(a)
kp_1, desc_1 = sift.detectAndCompute(image1, None)
image2 = cv2.imread(b)
kp_2, desc_2 = sift.detectAndCompute(image2, None)
matches = flann.knnMatch(desc_1, desc_2, k=2)
good_points = []
if good_points == 0:
continue
for m, n in matches:
if m.distance < 0.6*n.distance:
good_points.append(m)
number_keypoints = 0
if len(kp_1) >= len(kp_2):
number_keypoints = len(kp_1)
else:
number_keypoints = len(kp_2)
percentage_similarity = float(len(good_points)) / number_keypoints * 100
listOfSimilarities.append(str(int(percentage_similarity)))
listOfTitles2.append(b)
listOfTitles1.append(a)
countInner = 0
if a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countOuter += 1
zippedList = list(zip(listOfTitles1,listOfTitles2, listOfSimilarities))
print(zippedList)
dfObj = pd.DataFrame(zippedList, columns = ['Original', 'Title' , 'Similarity'])
dfObj.to_csv(r"/Downloads/images/DuplicateImages3.csv")
Answers:
I think significant performance improvements can obtained with simple changes:
- Firstly, since you are interested in comparing pairs of images, your loop can look like:
files = ... # preload all file names with glob
for a_idx in range(len(files)):
for b_idx in range(a_idx, len(files)): # notice loop here
image_1 = cv2.imread(files[a_idx])
image_2 = cv2.imread(files[b_idx])
This considers all pairs without repeating e.g. (a, b) && (b, a)
- Secondly, you don’t need to recompute the features for a when comparing for each b
for a_idx in range(len(files)):
image_1 = cv2.imread(files[a_idx])
kp_1, desc_1 = sift.detectAndCompute(image1, None) # never recoompute SIFT!
for b_idx in range(a_idx, len(files)):
image_2 = cv2.imread(files[b_idx])
kp_2, desc_2 = sift.detectAndCompute(image2, None)
- I would also check the image sizes. My guess is there are some really large ones that are slowing down your inner loop. Even all 60*60 == 3600 pairs should not take that long. If an image is really big, you could downsample it for efficiency.
I ran your existing implementation on my computer, on 100 images. That code took 6 hours and 31 minutes to run. Then I changed the implementation as I had suggested in my comment to compute sift.detectAndCompute once only for each image, cache the results and use the cached results in the comparisons. This reduced the execution time on my computer on the same 100 image from 6 hours 31 minutes to 6 minutes and 29 seconds. I don’t know if this will be fast enough for all of your images, but it is a significant reduction.
See my modified implementation below.
from __future__ import division
import cv2
import numpy as np
import glob
import pandas as pd
listOfTitles1 = []
listOfTitles2 = []
listOfSimilarities = []
# Sift and Flann
sift = cv2.xfeatures2d.SIFT_create()
index_params = dict(algorithm=0, trees=5)
search_params = dict()
flann = cv2.FlannBasedMatcher(index_params, search_params)
# Load all the images1
countInner = 0
countOuter = 1
folder = r"/Downloads/images/**/*"
folder = "SiftImages/*"
siftOut = {}
for a in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
image1 = cv2.imread(a)
kp_1, desc_1 = sift.detectAndCompute(image1, None)
siftOut[a]=(kp_1,desc_1)
for a in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
(kp_1,desc_1) = siftOut[a]
for b in glob.iglob(folder,recursive=True):
if not b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countInner += 1
print(countInner, "", countOuter)
if countInner <= countOuter:
continue
#### image1 = cv2.imread(a)
#### kp_1, desc_1 = sift.detectAndCompute(image1, None)
####
#### image2 = cv2.imread(b)
#### kp_2, desc_2 = sift.detectAndCompute(image2, None)
(kp_2,desc_2) = siftOut[b]
matches = flann.knnMatch(desc_1, desc_2, k=2)
good_points = []
if good_points == 0:
continue
for m, n in matches:
if m.distance < 0.6*n.distance:
good_points.append(m)
number_keypoints = 0
if len(kp_1) >= len(kp_2):
number_keypoints = len(kp_1)
else:
number_keypoints = len(kp_2)
percentage_similarity = float(len(good_points)) / number_keypoints * 100
listOfSimilarities.append(str(int(percentage_similarity)))
listOfTitles2.append(b)
listOfTitles1.append(a)
countInner = 0
if a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countOuter += 1
zippedList = list(zip(listOfTitles1,listOfTitles2, listOfSimilarities))
print(zippedList)
dfObj = pd.DataFrame(zippedList, columns = ['Original', 'Title' , 'Similarity'])
### dfObj.to_csv(r"/Downloads/images/DuplicateImages3.csv")
dfObj.to_csv(r"DuplicateImages3.2.csv")
I have a directory of images that contains many unidentified duplicates. My goal is to identify the duplicates. Because the duplicates have been cropped, resized, or converted to a different image format, they cannot be detected by comparing their hashes.
I wrote a script that successfully detects duplicates, but there is one major drawback: the script is slow. On a test-drive with a folder containing 60 items, it took five hours to run (this might also be a reflection of my increasingly buggy and slow computer). Since I have approximately 66,000 images in my directory, I estimate that it will take 229 days for the script to complete.
Can anyone suggest solutions? My research has revealed that you can free up memory by "releasing" the image stored in the variable as the loop completes, but all the information on how to do this seems to be written in C, not python. I was also thinking of trying to use orb instead of sift, but have concerns about its accuracy. Does anyone have advice on which of the two options would be better to pursue? Or a way to rewrite the script so it take less memory? Many thanks in advance.
from __future__ import division
import cv2
import numpy as np
import glob
import pandas as pd
listOfTitles1 = []
listOfTitles2 = []
listOfSimilarities = []
# Sift and Flann
sift = cv2.xfeatures2d.SIFT_create()
index_params = dict(algorithm=0, trees=5)
search_params = dict()
flann = cv2.FlannBasedMatcher(index_params, search_params)
# Load all the images1
countInner = 0
countOuter = 1
folder = r"/Downloads/images/**/*"
for a in glob.iglob(folder,recursive=True):
for b in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if not b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countInner += 1
print(countInner, "", countOuter)
if countInner <= countOuter:
continue
image1 = cv2.imread(a)
kp_1, desc_1 = sift.detectAndCompute(image1, None)
image2 = cv2.imread(b)
kp_2, desc_2 = sift.detectAndCompute(image2, None)
matches = flann.knnMatch(desc_1, desc_2, k=2)
good_points = []
if good_points == 0:
continue
for m, n in matches:
if m.distance < 0.6*n.distance:
good_points.append(m)
number_keypoints = 0
if len(kp_1) >= len(kp_2):
number_keypoints = len(kp_1)
else:
number_keypoints = len(kp_2)
percentage_similarity = float(len(good_points)) / number_keypoints * 100
listOfSimilarities.append(str(int(percentage_similarity)))
listOfTitles2.append(b)
listOfTitles1.append(a)
countInner = 0
if a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countOuter += 1
zippedList = list(zip(listOfTitles1,listOfTitles2, listOfSimilarities))
print(zippedList)
dfObj = pd.DataFrame(zippedList, columns = ['Original', 'Title' , 'Similarity'])
dfObj.to_csv(r"/Downloads/images/DuplicateImages3.csv")
I think significant performance improvements can obtained with simple changes:
- Firstly, since you are interested in comparing pairs of images, your loop can look like:
files = ... # preload all file names with glob
for a_idx in range(len(files)):
for b_idx in range(a_idx, len(files)): # notice loop here
image_1 = cv2.imread(files[a_idx])
image_2 = cv2.imread(files[b_idx])
This considers all pairs without repeating e.g. (a, b) && (b, a)
- Secondly, you don’t need to recompute the features for a when comparing for each b
for a_idx in range(len(files)):
image_1 = cv2.imread(files[a_idx])
kp_1, desc_1 = sift.detectAndCompute(image1, None) # never recoompute SIFT!
for b_idx in range(a_idx, len(files)):
image_2 = cv2.imread(files[b_idx])
kp_2, desc_2 = sift.detectAndCompute(image2, None)
- I would also check the image sizes. My guess is there are some really large ones that are slowing down your inner loop. Even all 60*60 == 3600 pairs should not take that long. If an image is really big, you could downsample it for efficiency.
I ran your existing implementation on my computer, on 100 images. That code took 6 hours and 31 minutes to run. Then I changed the implementation as I had suggested in my comment to compute sift.detectAndCompute once only for each image, cache the results and use the cached results in the comparisons. This reduced the execution time on my computer on the same 100 image from 6 hours 31 minutes to 6 minutes and 29 seconds. I don’t know if this will be fast enough for all of your images, but it is a significant reduction.
See my modified implementation below.
from __future__ import division
import cv2
import numpy as np
import glob
import pandas as pd
listOfTitles1 = []
listOfTitles2 = []
listOfSimilarities = []
# Sift and Flann
sift = cv2.xfeatures2d.SIFT_create()
index_params = dict(algorithm=0, trees=5)
search_params = dict()
flann = cv2.FlannBasedMatcher(index_params, search_params)
# Load all the images1
countInner = 0
countOuter = 1
folder = r"/Downloads/images/**/*"
folder = "SiftImages/*"
siftOut = {}
for a in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
image1 = cv2.imread(a)
kp_1, desc_1 = sift.detectAndCompute(image1, None)
siftOut[a]=(kp_1,desc_1)
for a in glob.iglob(folder,recursive=True):
if not a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
(kp_1,desc_1) = siftOut[a]
for b in glob.iglob(folder,recursive=True):
if not b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
continue
if b.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countInner += 1
print(countInner, "", countOuter)
if countInner <= countOuter:
continue
#### image1 = cv2.imread(a)
#### kp_1, desc_1 = sift.detectAndCompute(image1, None)
####
#### image2 = cv2.imread(b)
#### kp_2, desc_2 = sift.detectAndCompute(image2, None)
(kp_2,desc_2) = siftOut[b]
matches = flann.knnMatch(desc_1, desc_2, k=2)
good_points = []
if good_points == 0:
continue
for m, n in matches:
if m.distance < 0.6*n.distance:
good_points.append(m)
number_keypoints = 0
if len(kp_1) >= len(kp_2):
number_keypoints = len(kp_1)
else:
number_keypoints = len(kp_2)
percentage_similarity = float(len(good_points)) / number_keypoints * 100
listOfSimilarities.append(str(int(percentage_similarity)))
listOfTitles2.append(b)
listOfTitles1.append(a)
countInner = 0
if a.lower().endswith(('.jpg','.png','.tif','.tiff','.gif')):
countOuter += 1
zippedList = list(zip(listOfTitles1,listOfTitles2, listOfSimilarities))
print(zippedList)
dfObj = pd.DataFrame(zippedList, columns = ['Original', 'Title' , 'Similarity'])
### dfObj.to_csv(r"/Downloads/images/DuplicateImages3.csv")
dfObj.to_csv(r"DuplicateImages3.2.csv")