Automatic Web Image Annotation via Web-Scale Image Semantic Space Learning

Abstract. The correlation between keywords has been exploited to im-prove Automatic Image Annotation(AIA). Differing from the traditional lexicon or training data based keyword correlation estimation, we pro-pose using Web-scale image semantic space learning to explore the key-word correlation for automatic Web image annotation. Specifically, we use the Social Media Web site: Flickr as Web scale image semantic space to determine the annotation keyword correlation graph to smooth the annotation probability estimation. To further improve Web image annotation performance, we present a novel constraint piecewise penalty weighted regression model to estimate the semantics of the Web im-age from the corresponding associated text. We integrate the proposed approaches into our Web image annotation framework and conduct ex-periments on a real Web image data set. The experimental results show that both of our approaches can improve the annotation performance significantly.