Research On Image Fusion Methods Based On Statistical Modeling In Multiscale And Multidirectional Transform Domain

Image fusion is a process that two or more source images are fused into a new image by a certain algorithm. The fused image can preserve the useful information of the source image, and remove the redundant information. Usually, the fusion image has higher accuracy and robustness for observation, and better credibility and intelligibility, which is more beneficial to the effective, reliable and comprehensive image information post-processing for humans and computers. As an important branch of information fusion, image fusion technology is widely researched and applied in computer vision, artificial intelligence and image understanding. Usually, image fusion methods include the approaches of spatial domain and transform domain. The image fusion methods based on multi-scale and multi-directional transform can process the image information from low frequency to high frequency, and get a good fusion result and performance. At present, the research topic is a hot spot in the field of image fusion.The image fusion method based on multi-scale and multi-directional transform is widely used in the field of image fusion. It includes image fusion methods based on Pyramid transform, wavelet transform-and Contourlet transform. The analysis of multi-scale and multi-directional transform has obvious advantage in capturing the structure information of the edge and contour of the image. A reasonable statistical model establishing through the transform domain, can make the image fusion effect better by combining with the fusion rules which are set by the accurately estimating parameters of the model. Therefore, it has important theoretical and applied significance to study the method of image fusion based on the statistical modeling in multi-scale and multi-directional transform domain.Firstly, the development and the present situation of image fusion at home and abroad are introduced in this paper. Some classical image fusion methods based on space domain and transform domain are reviewed. The basic principle and framework of multi-scale and multi-directional image fusion are analyzed. The evaluation criteria and classification of image quality are summarized.Secondly, A multi-focus image fusion method based on statistical model in wavelet domain is proposed. Based on the Burt fusion rules, the image fusion is realized using wavelet transform and combining principal component analysis (PCA) and generalized Gaussian distribution (GGD) modeling. Using the methods of subjective evaluation and objective evaluation, compares and analyzes the experimental results with method of the weighted average, method based on PCA and method based on local energy. Experimental results show that the fusion images obtained by the proposed fusion algorithm have more abundant information and higher fidelity.Thirdly, a method of image fusion based on statistical modeling in Contourlet transform domain is analyzed. In this paper, based on Contourlet transform, the image fusion method combining Gaussian mixture model (GMM) with generalized Gaussian distribution (GGD) model is proposed. The high frequency subbands and low frequency subband in Contourlet transform domain are modeled by generalized Gaussian modeling and Gaussian mixture modeling respectively. According to the estimated parameters, image fusion is achieved based on improved Burt fusion rule and posterior probability weighted average respectively. The experiments are conducted by using the multi-focus image, the remote sensing image and the medical image respectively. The experimental results of this paper are compared and analyzed with the experimental methods based on wavelet transform under the same experimental conditions. Experimental results show that the proposed algorithm not only can enhance the spatial details of the fusion image, but also has more abundant information, and better visual effect and performance. And, the proposed fusion algorithm is suitable for a wide kind of source image fusions.Finally, the method of image fusion based on statistical modeling in Pyramid dual directional filter bank (PDTDFB) transform domain is proposed. The PDTDFB transform has better multi-scale and multi-directional analyzing ability. It also has more rich directional selection, more powerful translation invariance. Based on the transform, the Laplacian mixture model (LMM) is used to model the sub-bands of each direction, and then the fusion parameters are determined to obtain the fused image. The experiments are conducted by using the multi-focus image, the remote sensing image and the medical image. The experimental results of this paper are compared with the experimental methods based on wavelet transform under the same experimental conditions. The experimental results show that this method can effectively fuse source images of various types, and remarkably maintain the detail information.