Research On And Application Of Image Fusion Algorithms With Statistical Correlation Analysis And Visual Characteristics

Image information fusion is able to combine the different images of the same target or scene into a complete and accurate description of the same target or scene with software tricks. It has a wide range of applications in the field of medicine, remote sensing and military. A good image fusion method can lay a solid foundation for the subsequenct computer automated processing. This dissertation analyzes the research status quo of image fusion, and then does a deep research on the theory of various image fusion methods. Aiming at improving the image quality and benefiting the follow-up applications, this dissertation researches into noisy image fusion, the role that image quality evaluation plays in image fusion rules, multi-spectral and multi-temporal remote sensing image fusion, and weight construction based on learning algorithm. The main contributions are summarized as follows:(1) A fusion framework which contains multi-scale decomposition framework and Totoal Variation (TV) model is proposed for noisy image fusion. In the multi-scale decomposition framework, applying Principal Component Analysis (PCA) to accomplish the decomposition and reconstruction is presented. If one transform can be applied to multi-scale decomposition, it must have the ablity of transform and inverse transform. PCA can be applied not only in the decomposition but also to reconstruct images, so it is feasible to be used in multi-scale decomposition framework. The proposed framework solves the massive effect that occurs during the fusion process with TV model while overcomes the weakness of poor noisy suppression in the multi-scale framework. The experiments demonstrate that, with TV model, this new framework is able to reduce the noise of the fused image while maintain the spectral information during the fusion of multi-spectral image and panchromatic image; the proposed decomposition method based on PCA can avoid the massive effect appeared in the TV model fusion.(2) Image fusion rules based on image quality assessment are proposed. Fusion rules in multi-scale decomposition contain approximation images’ fusion rule and detail images’ fusion rule. Usually, when fusing the images from certain decomposition level, fusion rules ultilize the relevant information of current level’s images to accomplish the fusion. Obviously, as the increase of decomposition levels, the amount of information in the images gradually reduces. Therefore, it is proposed that using last level’s information of multi-scale decomposition to complete the fusion of current level’s approximation images. Corresponding Image quality assessment parameters are put into establishing the rules of approximation images and detail images. The experiments under different multi-scale decomposition frameworks show that the proposed rules are able to effectively extract information for fusion from approximation images and detail images.(3) For multi-spectral and multi-temporal remote sensing images, Multi-set Canonical Correlation Analysis (MCCA) is applied to approximation image fusion’s rule. MCCA is able to extract the similar information from group of objects. When applied to the fusion of approximation images, it can eliminate information with less correlation and at the same time improve the quality of fused image. In this method, wavelet transform is first used to acquire the approximation images of remote sensing images; then approximation images’ fusion is executed by MCCA and ’select max’ is considered as the detail images’ fusion rule; after inverse transform, the fusion of images withinclass is realized. Compared with other methods, the rule based on MCCA is able to handle many groups of images at the same time and fusion images acquired from MCCA have closer correlation with source images.(4) For training set, a method combined Kernel Generalized Canonical Correlation Analysis (KGCCA) with wiener filter and corresponding fast algorithm is presented, then it is applied to seek the similar images and construct weight for fusion. In this way, a new fusion method which can be used to image denoising is generated. In this method, the proposed KGCCA is first used to extract features from images. Furthermore, Kernel Half Generalized Canonical Correlation Analysis (KHGCCA) is presented based on Kernel Half Canonical Correlation Analysis (KHCCA). To reduce the computation cost in the kernel space and improve the efficiency of the algorithm, a proposed transform matrix is added into the original algorithm. After that, extracted features and wiener filter are used to construct the similarity function. In the end, the value of function is considered as the weight distributed to the corresponding image in the training set. With the weight, the denoising images can be acquired. In order to verify the recognition performance of extracted features with new algorithm, experiments are done in the face database as well as in the handwritten character database. The experiments show that, features extracted with new algorithm have better recognition performance than others. In the denoising experiments, the good results of restoration are further proofs of proper fusion weight. Comparisons of time consumption and performance of the algorithm indicates that the proposed fast algorithm method can improve efficiency while still obtain better fusion performance.