The Parameter-transfer Cluster Based On Gaussian Mixture Model

The Synthetic Aperture Radar (SAR) has ability of detection and investigation and so on for its all time and all weather. It makes use of pulse compression technique and synthetic aperture principle to get high resolutions on range and azimuth. So it has unique effect than real aperture radar in remote sensing field. The understanding and interpretation of SAR images belong to image processing which involves signal processing, machine learning and pattern recognition and so many subjects. Because of unique effect of SAR, the understanding of SAR image is paid more and more attention on both military and civil fields. The segmentation of SAR image which is one of the most key part of SAR image investigation is become more important. The cluster is very common way to deal with the image segmentation, so this paper mainly aims at the study of SAR image segmentation. The main contributions can be summarized as follows:(1) An EMBoost clustering based on spatial information for image segmentation is proposed. Compared with the traditional EM clustering algorithm, the EMBoost clustering algorithm can improve two problems that the sensitive result to initial value and the low precision. However, the local information is not considered in the EMBoost algorithm, which is useful to enhance the performance of the EMBoost algorithm, especially for image segmentation. Hence, we proposed a new approach that spatial information is brought into EMBoost clustering algorithm, which consisted of the adjacent pixels relative position and the neighbor texture distance, in order to improve the performance EMBoost clustering method..(2) An algorithm of the GMM parameter-transfer cluster based on spatial information is proposed in this paper, which is used on SAR image segmentation. The common technique of cluster is machine learning which has got significantly succeed. While most machine learning methods are based on that the training data and test data are come from same distribution and feature space. So when the data distribution is changed, most machine learning methods should study again from the very beginning and require re-collect lots of training data to train model. Re-collect data and label need expend rich price in real world. In additional, the expectation maximization algorithm was widely used for its simple and easy realize. While the EM algorithm itself is faulty with defects:sensitive to initial value which is lead to unstable result and miss segment. So the algorithm is proposed to improve these faults. (3) An cluster method based on block parameter transfer is proposed. We often meet large scale, ultra-high dimensional and complicated distribution data in real world. If we use traditional EM algorithm to deal with them, the segmentation effect and time are limited. So the image is divided into many blocks not stand on single sample in this paper. Then to estimate parameters and transfer them based on these blocks. Finally we get the purpose of image segmentation.This paper is supported by the National Natural Science Foundation of China (No.61003198,60970067,60803097), the National and Commissions Science Project of China (9140A07011810DZ0107), the Fundamental Research Funds for the Central Universities ((JY10000902038, JY10000902001, K50510020001and JY10000902045) and the Provincial Natural Science Foundation of Shaanxi of China (2009JQ8016)...