Compact Representation For Image Search

Content Based Image Retrieval is a classic computer vision problem. As a kind of multi-media search technique, it can help users to find the similar image in the huge database and obtain related information. It can also arrange database and remove du-plicated images. Image representation plays key role for image search system in terms of its time/memory efficiency and accuracy. However, existing image representation method cannot meet the demand of image search application. Therefore, to obtain ef-fective and accurate image representation is both meaningful and challenge.In this thesis, we study the topic of "Compact Representation for Image Search", about which three works are discussed. The first one is Sparse Coding(SC) based fea-ture aggregation. In this work local features of one image are sparse encoded and then pooled into one fixed-length vector, which is considered as the image representation. It is the first time such method is adopted for image search. The second one is Op-timized Product Quantization(OPQ). Product Quantization(PQ) is recently proposed method for Vector Quantization and Nearest Neighbor Search. Its key idea is decom-posing full space into subspace, which constrains codebook as the Cartesian product of sub-codebook of each subspace. We optimize space decomposition of PQ:with jointly optimized space decomposition and codebook, the performance of PQ is significantly improved. We use OPQ to further compress image representation. The third one is Product Sparse Coding(PSC). Coding efficiency is very important when it is used for image search, so we propose Product Sparse Coding to accelerate encoding procedure. This method also introduces subspace, through which codebook is constrained as sub-codebooks’ Cartesian product. By separating SC problem in full space into subproblems in subspace, we greatly reduce the encoding complexity.Various qualitative/quantitative experiments verified these works’ effectiveness and their positive affection for extracting image representation.