A graph theory-based online keywords model for image semantic extraction

Image captions and keywords are the semantic descriptions of the dominant visual content features in a targeted visual scene. Traditional image keywords extraction processes involves intensive data- and knowledge-level operations by using computer vision and machine learning techniques. However, recent studies have shown that the gap between pixel-level processing and the semantic definition of an image is difficult to bridge by counting only the visual features. In this paper, augmented image semantic information has been introduced through harnessing functions of online image search engines. A graphical model named as the "Head-words Relationship Network" (HWRN) has been devised for tackling the aforementioned problems. The proposed algorithm starts from retrieving online images of similarly visual features from the input image, the text content of their hosting webpages are then extracted, classified and analysed for semantic clues. The relationships of those "head-words" from relevant webpages can then be modelled and quantified using linguistic tools. Experiments on the prototype system have proven the effectiveness of this novel approach. Performance evaluation over benchmarking state-of-theart approaches has also shown satisfactory results and promising future applications.