Research On Fusion、Segmentation And Cooperation For Multisource Image Under Data Assimilation Framework

With the development of sensors and automatic data acquisition technology, acquiring multi-source image data has become easier. However, the information extraction of the data far falls behind the practical applications. Image fusion method can extract and integrate information from several images of different sensors and the fused image will benefit visual interpretation or the following computer-processing step. From the perspective of information theory, fused image contains more information than any of source images and suitable image fusion can establish a good foundation for subsequent image analysis phase.In image processing community, object recognition for specific application and extraction of intrinsic information by image analysis both are important tasks. Moreover, a good image segmentation result is the foundation of aforementioned tasks. Image segmentation algorithms can be divided into two subsets:the feature-based segmentation and the model-based segmentation. The research idea in this paper is two fold:on one hand, multi-source image fusion is regarded as a way to improve the feature, on the other hand improvement on classical MRF model is developed to create more accurate model, moreover the research takes image fusion and the image segmentation as a system to account, that makes the feature and model in image segmentation couple together. This point is the main difference between the work and other’s. The major contributions of this thesis are as follows:(1) Some classical image segmentation models and algorithms, such as MRF-based image segmentation, and the current research situation of image segmentation and fusion coupling are systematically summarized. Several image fusion methods’ influences on special image segmentation algorithm are also investigated, and the relationship between fused image and the segmentation effect is quantitatively analyzed.(2) The idea of data assimilation system in meteorology field is brought into image processing field, and an adaptive fusion framework is proposed. Under this framework, the different fusion algorithms can be used as the model operator and observation operator. The objective function is composed of a weighted sum of image measures and intelligent optimization algorithm is employed to obtain a proper image. This controllable fusion framework forms a solid foundation for the coordination of fusion and segmentation.(3) To deal with the randomicity and fuzzification in the image segmentation, an image segmentation algorithm based on fuzzy MRF model is expended into wavelet domain. The algorithm exploits adequately the ability of description fuzzification of fuzzy clustering algorithm and the ability of spatial information integration of MRF, and the multiscale merit of wavelet transformation. First, the multiresolution expression is obtained by wavelet decomposition. Then, from the coarsest resolution to the finest one, the MRF model and fuzzy clustering on each resolution is employed to obtain an initial segmentation result, and then the result is projected to the next scale as the priori information. At last, the result in the finest resolution is seen as final result. the experimental results using synthetic texture images, remote sensing images and medical images show better performance compared with FCM and MRF respectively both on visual effect and quantitative evaluation.(4) Traditional MRF takes pixels as processing unit, which is easily affected by image noise. A region-growing model under the framework of MRF is proposed for urban detection to reply this problem. In the framework, the basic processing unit is over-segmentation region. Firstly, this model obtains the initial seed points by texture analysis. Secondly, the over segmentation regions are got by Mean Shift algorithm and the regions that include seed points are set to be seed regions. Thirdly, it applies MRF to model these regions and determines the growing direction according to spatial neighborhood, uses the MAP to realize the area growing, which can get rid of the disadvantage of fixed growing direction. Finally, area threshold is assigned to control the growing process. The experiments using QuickBird and Ikonos images demonstrate that the model can effectively detect the urban area from the images.(5) A collaboration framework for image fusion and segmentation inspired by the data assimilation idea is provided. In the proposed strategy, the fusion rule and the parameters of fuzzy MRF model could be adjusted to obtain better multi-source image segmentation result. The multi-source images are used as the driven data of the data assimilation system, selectable fusion algorithms are used as the model operator and observation operator, the quantitative assessment index of segmentation is used as the objective function, and genetic particle swarm optimization algorithm is employed. Based on the framework, a semi-supervised image segmentation algorithm is proposed. Experimental results using complex texture images and remote sensing images show the good performance of the algorithm.