Nowadays face recognition and more generally, image recognition have many applications in the modern world and are widely used in our daily tasks. In this paper, we propose a novel distributed approximate nearest neighbor (ANN) method for real-time face recognition with a big data-set that involves a lot of classes. The proposed approach is based on using a clustering method to separate the data-set into different clusters, and specifying the importance of each cluster by defining cluster weights. Reference instances are selected from each cluster based on the cluster weights and by using a maximum likelihood approach. This process leads to a more informed selection of instances, and so enhances the performance of the algorithm. Experimental results confirm the efficiency of the proposed method and its out-performance in terms of accuracy and processing time.