Research And Application Of Semi-supervised Classification Methods For Large-scale Images

The Semi-Supervised Learning is an important method in the Machine Learning. Comparing with Unsupervised Learning and Supervised Learning, which takes advantage of the information from the labeled and unlabeled data and thus has more excellent classification results. Because of the advantage of needing smaller number of labeled data than supervised learning, the semi-supervised learning get more attention from people. Graph-based Semi-Supervised Classification method is an important method in Semi-supervised Learning, which has more advantages such as the framework is simpler to understand and easy to solve and has a solid theoretical foundation and so on. Graph-based Semi-Supervised image classification is an important application of Graph-based Semi-Supervised classification, but when we use it to classify the big size image, it couldn't work well because of some reasons such as the method needs to take a high graph and need a high Composition scale, huge Computing space.By researching these problems in the Graph-based Semi-Supervised image classification, our specific work on these problems in this paper are as follows:1. For the problem that the image data size is too large to build Graph and to lead higher complexity for classifying, we propose the semi-supervised image classification method based on data reduction in this paper. Firstly, this method cluster image based on local density. Then several samples are drawn from each class, and those samples are combined with all the labeled samples to constitute a reduced data subset. Next, this subset is classified by using semi-supervised classification method. The classification result of the original image is obtained according to a new clustering hypothesis method and we got the classification results finally. Based on Data Reduction, the semi-supervised classification method effectively reduces the size of the data to be classified. It could reduce the noise interference because that the classification map extracted from a subset of data is more suitable with information of each cluster. The experiments show that compared with AGR algorithm, the proposed method not only fits for large scale images, but also improves the classification accuracy, and reduces the complexity. It can be applied involving large scale images in more fields.2. In this paper, we present a new graph-based semi-supervised classification algorithm:minimum cost path label propagation (MCPLP). The method propagates label by the main path which is called minimum cost path between nodes and then we proved that the path is the best path to propagation label. MCPLP has the lower time complexity because it only propagates label along the minimum cost path and every data only need to be propagated once. We find there is connectivity problem comes from sparse graph existing in the proposed and other graph-based label propagation semi-supervised classification methods. This problem may make the label cannot be propagated to some nodes, which means some data cannot be classified. In order to solve this problem, we present a sparse symmetric graph construction method to improve the connectivity of the graph and a strategy of re-classify the data in the disconnected regions. Then all of the data can be classified. The analysis and experiments demonstrate that the MCPLP method not only possesses low classification time complexity, but also obtains high classification accuracy. The experiments for the large-scale image data show that the proposed method is fit for large-scale data classification.3. This paper also studied the semi-supervised classification method based on graph to hyperspectral remote sensing image. In this paper, we use MCPLP method and another comparison methods to experiment in some remote sensing images. The experiments of MCPLP on remote sensing image show that the MCPLP can be used in the hyperspectral remote sensing image classification.