I play a little with a semblance of image. I actually play with an image search system. Ideally, I want to create some kind of image index, which I can query to get similar images.
My current thought is to store some kind of ImageDescriptor in the index, and each descriptor can have different functions in it, for example. k-mean-cluster-centroids, histograms, ... And I have a simple calculation based on wight - each function has a distance function, and the result of this function is multiplied by wight and summed over all functions. The final amount is the distance from my image. Not sure if this is a good idea?
So, I started playing with histograms. I saved the histogram index, which is why I requested them for the distance between the histogram and the stored histograms of the index. This gives some resemblance, but in most cases is far from a good ideal.
Now I am playing with k-medium clustering. I have already implemented segmentation based on RGB distance (we will also try in Lab color mode). My index consists of a centroid vector (from clustering). Now I am doing only a minimal comparison between centroids. This gives better results, but also far from good.
My first question is: can I do something better with segments (clusters) than request a distance? How to include form information?
As a note, most of the images are images of everyday objects (different pencils, different glasses, different shoes, ...) and different textures on the background of the same color. There are no natural images, faces, trees, clouds, mountains, ...
Relations Zaharije
source
share