Several methods based on visual methods (BoVW, VLAD, ...) or recent deep leaning methods try to solve the CBIR problem. Bag of visual words (BoVW) is one of most module used for both classification and image recognition. But, even with the high performance of BoVW, the problem of retrieving the image by content is still a challenge in computer vision. In this paper, we propose an improvement on a bag of visual words by increasing the accuracy of the retrieved candidates. In addition, we reduce the signature construction time by exploiting the powerful of the approximate nearest neighbor algorithms (ANNs). Experimental results will be applied to widely data sets (UKB, Wang, Corel 10K) and with different descriptors (CMI, SURF).