Choosing a method for evaluating image similarity

Question

Algorithms

Good day to all, I am writing a program for comparing images. Ideally, there is a large collection of images and an arbitrary image you are looking for. Accordingly, it is necessary to determine the degree of similarity of the desired image and images from the collection. The formula I'm currently using has large errors. I have included it below.

I will not describe the whole algorithm, because long, I will go directly to the assessment of similarity. Suppose we have 2 images:
CI is an image from the collection (collection image).
SI - image for search (search image).
Both of them have a size of 50x50 and are presented either in the form of contours ( Canny edge detector ), or simply as a b/w image blurred with a Gaussian filter. Now I use a rather trivial formula:
where I1 are SI pixels and I2 are CI pixels.

But since the formula is linear, there are many inaccuracies in the query process. Hence the question, how to more accurately determine the degree of similarity of images?

Answer 1

At the moment for knowledge

Answer 2

habrahabr.ru/post/120562/

Answer 3

As published for you.
libccv.org/post/introducing-ccv-milestone/

Answer 4

ready-made solutions - kolmck.net/r_apps.htm (ImgSearch)