Compressive-signal annotation driven by a supervised topic-clustering BoF model

This paper presents a new Bag-of-Features model (BoF) to enhance the efficiency of automatic image annotation. Since the traditional BoF ignores the semantic of its vocabularies, it cannot be seen as descriptive representation of images in many image applications. To handle this critical limitation, firstly, we propose the RGB compressive texton. By using compressive sensing theory, the image can be compressed and its key information can be kept. Secondly, according to the topic of images, we extract RGB compressive texton from image of the same topic. Thirdly, the cluster algorithm is use to form clustering centers of each topic. Finally, using all topics cluster center to form new visual vocabularies of BoF model. Therefore each vocabulary has its semantics, which includes the topic information of images. We refer to such new BoF model as supervised topic-clustering BoF model. Experiments on automatic image annotation with a benchmark datasets Corel-5K show promising performance.