Compare two images

I seem to remember something called xdelta which does binary comparison of files. You could get the output from that to find out which files are most/least different.

xdelta is open source though so you would of course have to open your entire program.

A possibly slow option is to subtract one image from another. The one that results in an image closest to black would be the closest match.

You need to consider a lot of other transformations to get reasonably close to the usual meaning of "similar" wrt images.

Rotation, translation, cropping, scaling result in similar images most of the time, but a simple delta always classifies them as very different.

The various definitions of "similar" images can mean different things.

Finding the similarity of images is a very hard problem for computers, with a lot of active research. Right now, that research makes a good masters or Ph.D. thesis.

As a simple example, imagine the Cornell Box (good example) of a square room, a light on the ceiling, and two objects -- each can be a box or a sphere.

Would you consider the image 'similar' if the lighting levels are different? Are they similar if the color is different? Are they similar if an object is moved a little bit? Are they similar if the scene is rotated? Are they similar if the rendering algorithms are changed so they produce different inter-object lighting effects or adding soft shadows? Are they similar if the sphere is made out of different materials, such as refractive glass, reflective steel, or a matte texture? Are they similar if the image has been processed by a simple convolution filter? Are they similar if they are rendered as a pen-and-ink drawing?

More complicated definitions of 'similar' can involve the incredibly difficult problems of object recognition, feature tracking, feature extraction, and image registration, just to name a few.

Good luck on your project.

Hi, I'd recommend the opensource OpenCV library, which stands for Computer Vision.

It's a full-blown library which I used for a facial recognition app.

Site:
OpenCV

Google discussion (opencv answer):
Link

I admit it's probably to much for your needs, but it's capable of comparing two images using the following code:



I guess you should have no difficulty adjusting it for multiple images.

Greetings,

Starik

Quote:

cvMatchTemplate(...CV_TM_CCORR_NORMED);

Nope. It would fail for the examples I gave above, unless you could take steps to identify a suitable template region. It would also fail when the template region (even though it is color-normalized) and all parts of the compared image are substantially different in terms of rendered elements.

What might work with intel's library would be an attempt to back-project in order to regenerate the object, but that goes back to an object recognition system.

Without a precise definition of "similar", everything we mention is a wild guess anyway.

Besides, what does "most similar" to a reference image really mean? If we changed a matte black sphere to each of a horizontally striped texture, a vertically striped texture, and a diagonally striped texture, which is most similar?

Thanks to all for the replies.

I only know that the program must compare a image with a list of images and find which have the same shape but I don't know the complexity of the shape, only the next week I will have more information about the problem. Before the meeting I'd like to study the argument and with your advice I have a good point to start.

Sorry for my poor English. I must also study English.

Thanks again.
Lorenzo.