Research On The Multisource Image Fusion Methods

Multisource image fusion technology is widely applied in a variety of fields such as military surveillance, computer vision, medical diagnosis and remote sensing. In this dissertation, three levels of image fusion technology: pixel-level, feature-level and decision-level are studied, and some novel methods of analysis and processing are presented.The main research objective of pixel-level image fusion is to obtain a visual enhanced image. A new image multi-resolution fusion method based on statistical model is presented in typical multisource image fusion this thesis. The method results in restraining sensitivity to sensor noise since sensor noise item is introduced into the model. In multi-spectral image fusion, a new method is developed, which introduces correlative restriction into the statistical model. Using the method, the interrelated spatial information is well enhanced, and the spectral information of multi-spectral images is effectively kept.Feature integration and classification from multisource images are main issues in existing references. A new idea of feature-level fusion is presented in this thesis, whose concept is feature extraction is supported by multisource image fusion. Based on this idea, a new line extraction algorithm is developed by fusing the edge information of multisource image, where phase of edge is main fusion element and fusion rules are used to integrate different properties of multisource images. It can extract the line segments, which are not obtained by using single image or part images. Followed it, a road extraction algorithm from multi-spectral images is presented by using the line and spectral features.Decision-level image fusion is widely used. Dempster-Shafer theory of evidence is one of the main methods of decision fusion, but typical D-S theory is sensitive to highly conflict evidences. In this thesis, a new method based on the pretreatment mode is presented, where the basic probability assignments of conflict focal elements are partly transferred to the union of focal elements before using the Dempster combination rule, and the combination order are determined based on the conflicted measurement. Therefore D-S theory can deal with the cases with highly conflict evidences since conflict information is translated into unknown knowledge. Followed it, a classification algorithm for hyperspectral image data is presented, which providesbetter results than the method based on typical D-S theory.All algorithms presented are applied to real multisource image data. The experimental results show their validity and adaptability.