Application Reseasrch Of Natural Neighbor Graph Based Semi-supervised Learning For Image Retrievial Technology

With the rapid development of the Internet and the widespread use of electronic digital devices, massive multimedia data and image data have been created and disseminated. How to quickly and accurately obtain images from the massive data has become a hot research direction in the image retrieval field. Among them, the text-based image retrieval technology has become very mature. However, with a large number of new images quickly created and the rise of human labor, the traditional method via text markup no longer meets the needs of the development of image technology, and thus content-based image retrieval technology has great development potentials and practical values.Semi-supervised learning algorithms make full use of the information from labeled data and unlabeled data, and this characteristic fits the development requirements of content-based image retrieval technology. Semi-supervised learning methods have received more and more attention from researchers since it was first proposed. A variety of new algorithms have been proposed based on semi-supervised learning, and the semi-supervised learning algorithm based on graph is one of them. Manifold ranking is one of the representative semi-supervised learning algorithms based on graph, and its performance is closely dependent on the structure of the underlying graph structure. Therefore, it is critical to construct a graph which can reflect the structure of high-dimensional data space. The natural neighbor graph can adaptively be embedded into the intrinsic low-dimensional manifold structure of the high-dimensional data space and this capacity meets this requirement. Furthermore, it is not required to specify the parameter k during constructing the natural neighbor graph. Because of the problems of natural neighbors when processing datasets with outlier data points, in this paper the natural neighbor graphs are improved, then the improved natural neighbor graphs are applied to the image retrieval framework based on manifold ranking, and at last experiments show that the image retrieval algorithm based on natural neighbor graphs outperforms that based on k-NN.This master thesis mainly consists of the research work I have done during my graduate studies and the detailed contents include:First, the state of the art on image retrieval and semi-supervised learning algorithm from both the domestic and abroad is deeply studied. Besides, the natural neighbors and its potential application are also solidly investigated.Third, the construction of natural neighbor graph is improved. Because the traditional natural neighbor graph does not perform well on the dataset including natural outliers, this paper improves the natural neighbor graph, which achieves better results.Fourth, the improved natural neighbor graph is applied to the manifold ranking framework. Inspired by the k-nearest neighbor graph based image retrieval algorithm, this paper presents an image retrieval algorithm based on natural neighbor graph, and it is applied to the manifold ranking framework.Fifth, the effectiveness of natural neighbor graph based image retrieval algorithm is verified through experiments. In the experiments, the USPS handwritten digital image dataset and Cole dataset are used. For both datasets the image retrieval has been carried out.The experimental results showed that the effect of image retrieval based on natural neighbor graph is better than that based on k-nearest neighbor.