Demonstration of a similarity based image retrieval engine

Image of the base Your image

Image url :

This process computes a mixed similarity measure which takes into account the similarity in gray level, in structure and in local variations (details on the features). You can weight these components by modifying the following values:

V0 : V1 : V2 : V3 : V4 :

We are not using all the pixels of an image but only a subset of them. You can select the criteria used to define this subset as well as the data base. The base is made of 609 images randomly sampled from all the french tv channels plus 180 images of very different natures (portraits, textures, tv images, ...). For all the methods we extract only 250 interest points from each image. Maxima of the Haar transform Interest points in the contrast pyramid

Which kind of distance between characteristic's distribution do you want ? Correlation coefficient Bhattacharyya distance

Which kind of similarity combination do you want ? Weighted sum Ordered weighted sum

This process will show you the most similar images of our base, in decreasing order of the similarity index. How many images do you want to see ?

Do you want the details of the measurements to be printed Yes No

Refinement (for details, please click here). No refinement Maximum a posteriori Rank based refinement

In order to refine, I need the ranks of the best images (at most three)

and the ranks of the worst images (at most three)

Warning: The process and file format conversions performed by this system can be rather time consuming... It might take a couple of minutes for the result to be returned.