We aim at finding a suitable quantification algorithm to encode color invariant for indexing and retrieving images. To this end

an adaptive cluster network quantification algorithm for color invariants is proposed. By using a set of training images to train the color invariant vector

this algorithm can acquire a suitable color invariants vector having an adaptive number of members. In this paper

we discuss in detail how the threshold and the step in the algorithm influence the number of the vector members. After having done many experiments

we get a vector of 29 members for our training images when the threshold is 1 0 and the step is 0 3. In this setting

the vector is rather robust. Then we also apply the algorithm and another bench algorithm

named averaging quantification algorithm

to a content based image retrieval system. Experiments have been conducted on a database consisting of 1 126 images taken from different image databases. In order to evaluate and compare the querying results

an application specific software is developed. In the view of the correctness of the querying results

comparing the adaptive quantification algorithm with the averaging quantification algorithm

we find the former is superior to the latter by 4%. In the view of the time complexity

although the former takes a long time to train quantification vector and to acquire a lookup table for the image database

it is much superior to the latter when retrieval proceeds. Finally a conclusion is obviously obtained that adaptive quantificaion algorithm is an excellent quantificaion algorithm for color invariants.