A Hypergraph-based Semantic Information Fusion Method For Image Scene Classification

Along with the Internet technology and information industry development, and emergence of various image share website, now we have entered an era consisting of image. The image is an intuitive resource and plays an important role in daily life.Image classification has been important research field in both computer vision and machine learning. Traditional image classification method only utilize the image vision feature, but the gap between the vision feature and semantic meaning the image represents is huge, so that the classification performance is low. There are not so much researches on how to extract semantic information of image directly from the vision feature till now. The tags related to image are the most direct descriptions of the image. And researchers have tried to bring the tag information into the image classification task. With the emergence of Filckr and Facebook, it’s easy to fetch images attaching with tags. Hence, research how to fuse the tag information in the task of image classification is urgent. Researchers have employed multi-modal learning, multi-task learning and transfer learning methods to bring the image associated text information into image classification tasks. Experiments conducted have shown that it is useful to bring the image associated text into the image classification.As the higher order can be hold in hypergraph, hypergraph learning methods also can be used in image scene classification. The relationship among hyperedges have not been considered in the existing image scene classification methods based on hypergraph. In this thesis, we propose a image scene classification method based on hyperedge correlation. We extend the traditional hypergraph learning method, through adding the hyperedge correlation in the hypergraph learning. The correlation among hyperedges are measured by both image vision information and associated tag information. In order to validate the effectiveness of the proposed method, experiments have been conducted on LabelMe and UIUC datasets. We analyze the influence of different parameters to the performance of the method, and compare with other methods introduced tag information into image classification. The validity of the extended hypergraph learning method is verified by experiments.