Research And Implementation Of Inverted Index For Large-scale Visual Search

With the ever increasing using of internet and decreasing cost of mobile digitalcameras, digital images are popular in people's daily life and visual search is becomingto the urgent requirement. In the field of visual search, state-of-the-art techniquesemploy Bag of Words (BoW) based inverted indexing and achieve considerableperformance. However, one fundamental while unexploited challenge is to maintain thelarge-scale inverted indexing subject to the limited hardware resources with a singleserver, suffering from the search inefficiency. In order to solve the problem within largescale visual search scenario, this work focuses on the distributed visual search andproposes valuable algorithms. This work is important for the application of visualsearch.With the research of visual characteristic projects, first, to overcome the problemof inaccuracy in image collection allocating and selecting for image is lack of semanticmeanings in BoW based representation, we studied Latent Dirichlet Allocation (LDA)to discover the latent concepts as an intermediate semantic representation over a largescale image database, proposed a novel semantic related image allocation and collectionselection method for large-scale distributed visual search. Second, to tackle the problemof load unbalancing between servers, we exploit the concurrence statistics of visualwords to distribute the visual words of a given query as uniformly as possible, proposeda novel co-occurrence based distributed vocabulary indexing architecture. Experimentsshowed the advantage of our proposed approaches. Third, we implemented amobile-based distributed visual search system over10million image collections andfurther verified our proposed methods.