Semi - Supervised Image Classification Based On

Image classification is an important foundation of image analysis, understanding and recognition of the target, occupies an important position in the image project. There are many ways to solve this problem. The graph-based semi-supervised learning has attracted more and more attention, which uses labeled and unlabeled to be classified, the objective function is convex function, and optimization easier. It is more efficient than supervised learning and unsupervised learning. Linear neighborhood propagation (LNP) is a graph-based semi-supervised learning algorithm, and using in classification problems. The following problems in practical applications:When creating graphical models in pixels, the image is too large will lead to a high computational complexity, and the number of neighbors selection will directly affect the calculation of the weight matrix. In addition, in the case of a small amount of labeled, target’s degree of difficulty, noise and texture will easy effect the image classification. Reduce the number of vertices, adapt to select the number of neighborhood and effectively remove noise or texture become necessary to help to improve the learner’s correct classification.To solve these problems exist in semi-supervised learning based on graph, we propose two algorithms for image classification on the basis of semi-supervised learning based on graph. The specific works of this paper are as follows:1) Introduced the concept of image classification and several common classification methods, the basic theory of semi-supervised learning and the state of the study in abroad. Then focused on theoretical foundation, the main method and the current research of the graph-based semi-supervised learning. What’s more, analyzed its defects and difficulties. Finally, the application of these methods are described. Then focused on introduction the linear neighborhood propagation.2) For images, LNP has high computational complexity for large images. Furthermore, if the number of neighbors is unsuitable, the classification results will be inaccurate. Therefore, a locally clustering based adaptive classification algorithm for LNP is proposed in this paper. The method improves the LNP classification algorithm in two aspects. Firstly, quick shift is used for local clustering to get the point-clusters. Point-clusters replace pixels as the graph model nodes. Then the size of matrix is reduced. Secondly, the relationship between the geodesic distance and the Euclidean distance is used to dynamically determine the number of neighbors for each point. The experimental results show that better classification results are obtained by the proposed method and the running time is largely reduced.3) When using the LNP algorithm, if the image contains more noise or texture, the number of vertices of the graph and calculating the similarity matrix become very difficult to solve. On the basis of previous paper, this article proposed an algorithm for LNP image classification based on image decomposition and watershed. Using the image decomposition to remove noise or most textures, to be making similarity matrix calculation more accurate. Meanwhile watershed algorithm reduce the number of vertices greatly in building graph. The computational complexity is to be reduced and the accuracy of classification is improved. Through comparative analysis of the experiment, the proposed method has a better performance on the classification.