Abstract: The fast approximate Kmeans algorithm (FAKM) was proposed to solve the limitations of traditional Kmeans algorithm in the large scale database. Based on the approximate Kmeans algorithm (AKM), FAKM classifies the cluster centers according to cluster results. This new algorithm filters out the cluster centers with few samples, and makes good use of those with intensive and stable samples, and thus the number of samples and clusters will reduce in each iteration. Accordingly it can improve the speed of this algorithm and refine the cluster result. Several experimental results in image retrieval system are presented to demonstrate its average advantage over Kmeans and AKM in the clustering time, retrieval time and the robustness capability of retrieval accuracy.

Key words: fast clustering, approximate nearest neighbor, image retrieval, large scale database