Hi Corona688,

The whole idea of compression for image retrieval is that it is a measure of entropy. If the data is completely random, then S(12) should be result in pretty much the same file size as S(1) and S(2), therefore D_comp -> 0 (i.e. D_comp converges to zero).

Your comments about a one DN value difference in a pixel may as well be a 255 difference is touched upon in the paper.

The paper does specifically describe the method in that it uses gzip 1.2.4 and that it uses the LZ77 lossless compression. It also states that the images are raw images, it doesn't elaborate further. I have taken that to mean that they are header-less images, so RGB images with 8bit colour depth converted using image magick.

I agree that cat isn't appropriate, however as a quick test I thought it would still work. It mentions that the images are combined 'as juxtaposition of pixel arrays I1 and I2'. I have taken this to mean appending the two images side by side. So I have tried this, and also one image followed by another.

Since there are lots of papers on this technique, unfortunately most referring back to this one for how it is implemented, I am sure the theory is sound. Of course I am aiming to test that for a few other situations, hence the research. I just think that I am missing some implementation subtly that isn't discussed in the paper. I am still convinced that it is the header.

If someone could explain why two files compressed individually would have a total file size less than if both of them were compressed together, then that would be great. As far as I am aware, that should not happen, even if they are completely random pieces of data.

Thank you