With the rapid development of the technologies and great advances in computer hardware, digital devices and communication transmission, we are entering the age of information. The image resources grow explosively on the Internet due to the improvements of multimedia and web techniques. As an important form of multimedia data, image usually contains massive information. How to represent and analyze image data is very important for understanding of image content. Moreover, due to the explosive growth of web images, the ability of managing huge image data and conducting fast image search in large-scale databases has also become a more and more important research issue, which is of great significance to the modern world. This paper employs graph theory, which is widely used in machine learning methods, to model the intrinsic structure of images and applies it to the process of image content analysis. We conduct intensive research on image latent topic discovery, compact feature description and fast image search on large database. The main contributions of this dissertation include the following issues: (1) We propose a novel latent topic model with dual local consistency for image analysis. Based on the Bag-of-Words representation for images, the latent topic models are usually used to narrow the gap between the low-level image features and high-level semantic information in different tasks. In order to overcome the shortcoming of traditional topic models in the ability of topic description, we incorporate two different local constraints in topic discovering. We first construct a l1<上标!>-graph to model the sparse neighborhood structure of images, and then combine it with word co-occurrence information for topic learning to obtain more accurate probabilistic distribution of images in the latent semantic space. Our model achieves improvements in image clustering task compared with traditional models. (2) We propose a novel spectral hashing method with semantically consistent graph for image compact coding and fast retrieval. Hashing methods is purposely designed to speed up the Approximate Nearest Neighbor (ANN) search. Baded on the Spectral Hashing (SH) method, our approach propose a novel similarity measure for graph construction between images. Then, we employ the image tags provided by the users to learn the graph structure. Finally, we can obtain more effective binary codes of images based on the optimized graph and get higher precis...
修改评论