Fast Image Retrieval Using Classified Vector Quantization of Feature Space

Abstract

Based on the mean and variance values of a feature vector, we classify the feature vectors into four classes, obtaining four-class feature spaces. Next, the feature space of each class is vector quantized to obtain several sub-spaces (or clusters), each cluster being represented by a codeword. For each class, we sort the feature vectors in each cluster in the ascending order of their mean values. For any given query image, we first compute its feature vector and then find its class based on the mean and variance values. The online image retrieval for a given query image is then progressively performed in the first several nearest clusters (from the nearest cluster to farther clusters) of this class based on a fast search method named updating enhanced equal-average equal-variance K nearest neighbor search (UEEEKNNS) to find the first several nearest neighbors of the query feature vector as soon as possible.