We focus on the problem of large-scale near duplicate image retrieval. Recent studies have shown that local image features, often referred to as key points, are effective for near duplicate image retrieval. The most popular approach for key point based image matching is the clustering-based bagof-words model. It maps each key point to a visual word in a code-book that is constructed by a clustering algorithm, and represents each image by a histogram of visual words. Despite its success, there are two main shortcomings of the clustering-based bag-of-words model: (i) it is computationally expensive to cluster millions of key points into thousands of visual words; (ii) there is no theoretical analysis on the performance of the bag-of-words model. We propose a new scheme for key point quantization that addresses these shortcomings. Instead of clustering, the proposed scheme quantizes each key point into a binary vector using a collection of randomly generated hyper-spheres, and a bag-ofwords model is constructed based on such randomized quantization. Our theoretical analysis shows that the resulting image similarity provides an upper bound for the similarity based on the optimal partial matching between two sets of key points. Empirical study on a database of 100, 000 images shows that the proposed scheme is not only more efficient but also more effective than the clustering-based approach for near duplicate image retrieval.

Local image features, such as SIFT descriptors, have been shown to be effective for content-based image retrieval (CBIR). In order to achieve efficient image retrieval using local features, most existing approaches represent an image by a bag-of-words model in which every local feature is quantized into a visual word. Given the bag-of-words representation for images, a text search engine is then used to efficiently find the matched images for a given query. The main drawback with these approaches is that the two key steps, i.e., key point quantization and image matching, are separated, leading to sub-optimal performance in image retrieval. In this work, we present a statistical framework for large-scale image retrieval that unifies key point quantization and image matching by introducing kernel density function. The key ideas of the proposed framework are (a) each image is represented by a kernel density function from which the observed key points are sampled, and (b) the similarity of a gallery image to a query image is estimated as the likelihood of generating the key points in the query image by the kernel density function of the gallery image. We present efficient algorithms for kernel density estimation as well as for effective image matching. Experiments with large-scale image retrieval confirm that the proposed method is not only more effective but also more efficient than the state-of-the-art approaches in identifying visually similar images for given queries from large image databases.

Bag-of-words model is one of the most widely used methods in the recent studies of multimedia data retrieval. The key idea of the bag-of-words model is to quantize the bag of local features, for example SIFT, to a histogram of visual words and then standard information retrieval technologies developed from text retrieval can be applied directly. Despite its success, one problem of the bag-of-words model is that the two key steps, i.e., feature quantization and retrieval, are separated. In other words, the step of generating bag-of-words representation is not optimized for the step of retrieval which often leads to a sub-optimal performance. In this paper we propose a statistical framework for large-scale near-duplication image retrieval which unifies the two steps by introducing kernel density function. The central idea of the proposed method is to represent each image by a kernel density function and the similarity between the query image and a database image is then estimated as the query likelihood. In order to make the proposed method applicable to large-scale data sets, we have developed efficient algorithms for both estimating the density function of each image and computing the query likelihood. Our empirical studies confirm that the proposed method is not only more effective but also more efficient than the bag-of-words model. W. Tong (B) Language Technologies Institute,

No abstract