Similarity Search Methods For Image Retrieval

With the rapid development of multimedia technology and thewidespread popularity of digital capture devices, the number of images hasgrown explosively. Although the rapid growth of the images brings much richand convenient information, it is difficult to search the image of interestefficiently in the large-scale image database. Therefore, how to organize thelarge-scale image database effectively and search the images of interestaccurately, is an urgent problem in the field of image retrieval.To solve this problem, the similarity search methods are investigated inthe field of image retrieval. These methods are described in the level ofsimilarity, spatial similarity and region similarity, and summarized as follows:1. In the level of similarity, we proposed a retrieval method usingordered quantization. This method introduces the ordered quantization toincrease the distinction among the quantized feature descriptors. Thus, thefeature point correspondences can be determined by the quantized featuredescriptors, and they are used to measure the similarity between query image and database image. To compute the similarity efficiently, amulti-dimensional inverted index is proposed to record the number of featurepoint correspondences, and then approximate RANSAC is investigated toestimate the spatial correspondences of feature points between query imageand candidate images returned from the multi-dimensional inverted index.Experimental results demonstrate that our method improves the retrievalefficiency while ensuring the retrieval accuracy in the content-based imageretrieval.2. In the level of spatial similarity, we proposed a retrieval method usingscale invariant visual phrases. This method generates the co-occurringfeatures in the scale invariant feature detection, and quantizes theco-occurring features into scale invariant visual phrases (SIVPs). Because theSIVPs capture translation, rotation and scale invariance, they are employed todetermine the spatial correspondences between query image and databaseimage. To compute the spatial correspondences efficiently, the SIVPs areintroduced into the inverted index to record the number of spatialcorrespondences, and then orientation consistency algorithm is investigatedto verify the spatial correspondences between query image and candidateimages returned from the inverted index. Experimental results demonstratethat our method improves the retrieval accuracy while increasing the retrieval efficiency.3. In the level of region similarity, we propose a retrieval method usingadaptive rectangular windows. This method represents the local regions ofdatabase images by adaptive rectangular windows, which can not only reflectthe true distribution of local regions, but also increase the similarity withquery object region. Each window is then represented by an independentwindow vector, which is used to compute the similarity with the query objectvector. To ensure the real-time similarity search, this method constructs aninverted index based on window vectors, and introduces the informationentropy of window vectors into the inverted index, which refines the windowvectors and accelerates the searching process. Experimental resultsdemonstrate that our method improves the retrieval accuracy while ensuringthe retrieval efficiency.