A Framework of Large-Scale and Real-Time Image Annotation System

In this paper, we propose a novel framework of large-scale and real-time image annotation system. The large-scale image set is constructed based on current web image search engines and re-ranking algorithm. Various global and local features are employed for representing images with parallel extraction mechanism for the real-time requirements. At training stage, the distance between class centers in image set are calculated, on which numbers of local neighbor areas are formed. In each local neighbor area, a bi-coded genetic algorithm is employed to select optimal feature subsets and corresponding optimal weights for every one vs. one SVM classifiers. At annotation stage, after finding the nearest class center for an unlabeled image, a set of pre-trained SVMs in local neighbor area are used to vote and obtain the final annotation. All the above strategies guarantee the annotation precision while shorten the annotation time of system.