Need a (simple) algorithm to tell if 2 images differ

Question

imneo OP

Created Mar ’23

Replies 3

Boosts 0

Views 1.9k

Participants 3

I have a need to tell if 2 images are different by unimportant details.

For example, one image is only slightly different than another because:

*) It is a little stretched *) or it has 99% portion of another (only missing some pixels on borders)

I wonder if there is an efficient algorithm to achieve my goal. Any suggestions are welcome.

Answered by Claude31 in 746698022

are different by unimportant details.

Your criteria are a bit fuzzy,

if only border missing, nearly all pixels are the same

In this case, such algorithm could work: https://stackoverflow.com/questions/6488732/how-does-one-compare-one-image-to-another-to-see-if-they-are-similar-by-a-certai

but if stretched, all pixels will be different

This could do it: https://stackoverflow.com/questions/71615277/image-similarity-in-swift

I would also try to compute a similarity distance:

split the images in small squares (20*20 for instance), which would make a few hundreds square areas
compute the average color of each area
Set thresholds: similarity is the % of squares which average color differs less than 10% different; if similarity > 90% then images are "similar": you set 10% and 90% to suit best your need
a more sophisticated way would be to compare pixels histograms of the squares: https://developer.apple.com/forums/thread/725550?answerId=746000022#746000022

Hope that helps

Boost

Answer 1

Claude31 OP

Mar ’23

Accepted Answer

are different by unimportant details.

Your criteria are a bit fuzzy,

if only border missing, nearly all pixels are the same

In this case, such algorithm could work: https://stackoverflow.com/questions/6488732/how-does-one-compare-one-image-to-another-to-see-if-they-are-similar-by-a-certai

but if stretched, all pixels will be different

This could do it: https://stackoverflow.com/questions/71615277/image-similarity-in-swift

I would also try to compute a similarity distance:

split the images in small squares (20*20 for instance), which would make a few hundreds square areas
compute the average color of each area
Set thresholds: similarity is the % of squares which average color differs less than 10% different; if similarity > 90% then images are "similar": you set 10% and 90% to suit best your need
a more sophisticated way would be to compare pixels histograms of the squares: https://developer.apple.com/forums/thread/725550?answerId=746000022#746000022

Hope that helps

0

Answer 2

endecotp OP

Mar ’23

If I were you, I'd start by checking if OpenCV has something that you can use.

If you google "OpenCV image similarity" you'll find some StackOverflow posts with helpful advice.

I fear it is not as simple as you might have hoped.

1

Answer 3

endecotp OP

Mar ’23

Hi Imneo,

Note that if you post comments on this forum, it doesn't send notification email like it does for replies. So I suggest always posting replies.

I want to do this in Objective-C. Does anyone have experience on integrating OpenCV into Xcode project?

That's easier than integrating with Swift. Use an objective-C++ (.mm) file to bridge.

There is documentation about building for iOS here:

https://docs.opencv.org/4.x/d5/da3/tutorial_ios_install.html

I've not done that - I've only used OpenCV server-side - but it looks simple enough.

1