I don't know exactly but some of my previous room mates used to work by Gray Scale and other image factors in matlab.

You can make a matrix from different factors and then use guided neural network to find the most similar.

Not sure but it might be path.

The question is not clear but for similarity, you can use peak signal to noise ratio (PSNR) or non-correlation.

Thanks for your answers, but I am interested in the images, which are NOT similar. All images are snapshot of 50-100 computers display and we want to find the snapshots which are different from the other. I did not try PSNR yet, but I will give it try.

Any other suggestion?

Dear Darios,

Fist of all, you have to define your similarity function: what do you consider similar and what dissimilar? Are you thinking about pictures having the same gamut of colours, say greens and blues, and you want to take the red and yellow ones out? Are you thinking about their content, like you want to get your cat pictures out of your dog pictures?

As a general method, consider normalizing your pictures (uniform size, maybe convert to gray scale, normalize histograms, whatever) and then simply take euclidean distance: take the difference of pixel values, pixel-wise, add their squares and, if you will, take the root of the sum.

Other methods that may work are: PCA projection onto the first three PCs, dendrogram clustering or Kohonen maps.

Hope this helps.

Cheers,

Fernando

As a second thought, I guess you want to preserve colour, so instead of working in RGB I would suggest translating your image to a colour space that describes colour in a more psychological way, like HSV or HSL (see link attached).

For us, humans, a difference in saturation or lightness is not a big deal, but a difference in hue is a big difference. So, convert your images to HSV, normalize their size and take the euclidean distance of your hue (H) component.

Cheers,

Fernando

https://en.wikipedia.org/wiki/HSL_color_space

Dear Fernando,

thanks for your answer. I will explain more about this challenge. Imagine there are 500 students working on same program, but 10 students working not only on the same program but also looking at a browser. You have only snapshots images of their displays (2 shot pr. Minute). I want to find those 10 snapshots among more then 1000 of other images. The images are of *.jpg type but can be converted to gray tones (less suitable). So yes, I am thinking about their content. You suggestion regarding normalizing histograms and analyzing pixel values sounds very good. And your second method HSV and HSL are worth to try. Thanks

Dear Darios,

I believe the solution to your answer might need to be different from normalized histogram type features for the whole image. These are generally a good first shot for homogeneous texture type image description.

Since you are looking to identify content, you would need to define WHAT/WHICH content you are looking for. Then you can identify which local feature and search method to use to identify it. Once this is in place algorithmically I believe you can break down your problem into an object detection and an object identification step. I have no practical experience with these, but a good starting point might be recent publications in IJCV.

Also, why do you need to take images to perform this task? Can you not identify/detect the active windows for each user?

Dear Tûnde

Thank you very much for your answer.

I am interested to content of images, but I don’t know what it is. The only thing I know is that all images should be almost equal. If there are some images that are not similar to the other so I need to point them out to find out why. I cannot use pattern recognition because I do not know what I am looking for.

Why do I need to take images to perform this task?

Because I cannot get active windows of users computers, which are private computers in form of BYOD. I have only 1000-15000 images of their desktops.

I have also another challenge with pattern recognition in which I use matlab’s NCC to find out the best matches. But NCC works fine with gray color images. Do you know how to use NCC on color images?

Thanks!