Color Gravity Descriptor

Currently we are working on our new color descriptor for content-based retrieval of 2D static images (CBIR). The descriptor is based entirely on the color information and the positions of color centroids within the image plane. It is scale and rotation invariant and produces retrieval results that are better than the known MPEG-7 image descriptors. We have indications that the descriptor performs better than any other descriptor that is entirely based on color and no other structural-scene information, such as points of interest or texture. Our descriptor was named Color Gravity Descriptor (CGD) as the method involved centroids of color masses.
As the algorithm involves a couple of heuristics we are currently working on optimisations, looking for optimum conditions and possible limitations. We are planning to give it a try at IEEE ICIP2014…
CGD, in summary, applies color reduction, and for each color in the indexed image it calculates the centroid, its position (bounding box) within the image and its distribution (sparsity). It then forms a graph-like construct based on the centroids and connects these centroids using weights produced by the position/sparsity of the colors, thus forming the image signature.
The signature is N2 in length, N being the number of colors after the color reduction.
Extensive tests using the Wang database showed very promising retrieval results.
Some indicative retrieval results follow.

Leave a Comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.