Technique of Large-scale Image Set Construction Based on Web Image Searching Engine

Large-scale image training set is the precondition of large numbers of various images´ semantic annotation. However, due to the absence of image content, state-of-the-art text-based Web image searching engine´s results can´t serve as image set directly. In this paper, we propose a novel framework for large-scale image set construction, which is based on re-ranking current text-based Web image searching enginepsilas results. For a particular concept to be included in future image set, a genetic feature selection algorithm is performed to obtain optimal features and relevant optimal weights based on the results of Web image searching engine and users´ relevance feedback. With the optimal feature set and optimal weights, the distance between image in original searching results and positive or negative instances users provided is considered to be the main factor of rank score. After re-ranking thousands of concepts and obtaining numbers of images ranked on top for each concept, large-scale image set can be constructed.