Study Of Incomplete Image Data Classification Technology Using Tensor Analysis

In many practical problems, the so-called incomplete data exists in many lost image data or noise polluted data. These data generally exists in biomedical images, remote sensing images or computer vision applications. It is strong demand to solve these problems the practical application. Currently, the research trend in incomplete data focuses on reconstruction and rehabilitation, while classifications study is relatively a new research direction.Due to the lack of to powerful tools to extraction and describe the feature of incomplete data, development of incomplete data classification research has been delay. Late 1990s, a new method Tensor Voting which is based on the data described in tensor got further developed. It is a new structure feature extraction algorithm which matches with human visual perception characteristics. The sparse data is coded into tensor, and then extract spatial structure through the nonlinear tensor voting and corresponding feature extraction algorithm. That is to extract complete structure from incomplete data. Taking into account of the fact kernel learning method, especially support vector machines (SVM), is based on statistical learning theory as a basis for pattern classification algorithm, which possesses good Computational Efficiency, Robustness, Statistical Stability, and other characteristics, this method has obvious advantages in image classification. This dissertation integrates images tensor analysis with the study of kernel methods, and establishes a new kernel methods based on tensor analysis, and classifies the image on incomplete data sets.The work of this research focuses mainly on the classification issues of incomplete image datum. First, the dissertation summarizes tensor analysis and classification approaches. Then, the principle and characteristics of SVM is introduced. Finally, it elaborates image processing methods about tensor representation, tensor voting and feature extraction, and presents some improved tensor voting algorithm for different applications subsequently. On the basis of above work, this dissertation proposes a novel classification algorithm of the incomplete data, which takes tensor characteristic and gray characteristic as input features of SVM to study and classify. Then, the real satellite remote-sensing images blocked by clouds are utilized on the research to validate the feasibility of this algorithm, which is compared with the minimum distance, Maximum Likelihood and BP neural network. Experimental results show that this method can achieve fairly good classification effectively.