Research On Pixel-Level Image Fusion Algorithms

Image fusion is a part of information fusion. It has attracted more and more attention and has gotten rapid development with the improvement of sensor technology and the widely application of various sensors. Fused images with higher resolution and understandability by combing complementary information of source images, which make up for a single sensor imaging limitations and differences in spatial or spectral resolution. The description of the scene in fused images can be more comprehensive and more accurate that is suitable for human and machine perception or for further image processing tasks.The fusion algorithms of the thesis are focused on pixel level. Multifocus images and multisource images are selected as main research targets in this thesis. Multisource image fusion composes of the research on the image fusion of infrared image and visible image and the research of multispectral image and panchromatic image.The main contributions of the dissertation are summarized as follows:(1) A novel image fusion algorithm based on the combination of nonsubsampled Contourlet transform(NSCT) and spatial frequency is proposed. The original images are decomposed by NSCT into low-pass subbands and a series of high-pass subbands at different levels. In order to obtain clearer contour and more details, spatial frequency is used to select adapted coefficients of low-pass and high-pass subbands. The high-pass subbands preserve the details of the original images, so we use the spatial frequency selection principles for the subbands coefficients that can enhance the robustness of the algorithm.(2) Two algorithms for fusion of multifocus images are proposed. One method is to use a rule of Difference images between PM model and Gaussian filter(DPG), that is combined DGP with small windows accumulation method. The other one is DPG-SML method, which combines DPG and Sum-Modified-Laplacian algorithms. PM model can encourage smoothing at homogeneous region and simultaneously preserve its edges. In order to obtain difference coefficients of an input image, a difference coefficients map between PM model and Gaussian filter is produced. With the aim of improving the robustness of the fusion algorithm, the decision maps are constructed by accumulated values over a small window and SML, respectively. Additionally, the fused image is composed of original pixels which are chosen directly from the corresponding sources in terms of a selection rule of focus measures that avoids being contaminated by the blurred regions of the source images.(3) A fusion scheme to select adaptively weighted coefficients for the image fusion of infrared and visible images is proposed. The method is aimed at preserving high spatial detail information and showing infrared targets clearly in the resulting fused image. This scheme can be determined by the histogram of the infrared image that enhances segmentation small infrared targets and preserves spatial detail information from the infrared and visible images. Furthermore, A background modeling method of segmentation based on histogram is proposed. Moving object regions are identified by comparison of the current image with the robust background model that is built from an infrared image sequence. The shceme is suitable for image fusion of infrared and visible images and is used to mark the moving targets in special color that is useful for visual surveillance.(4) The proposed fusion scheme based on the IHS color space for the image fusion of multispectral image and panchromatic image. The multispectral image is composed of multiple channels, so the fusion method need to achieve in color space. In this thesis, the intensity component of fused image is obtained according to adaptively weighted algorithm to combine intensity component of multispectral image and panchromatic image in the IHS space. An automatic color equalization algorithm is adopted for enhancing resulting image after converting into the RGB space. The multispectral has good spectrum and the panchromatic image has high spatial resolution and geometric accuracy. The multispectral image is merged with the panchromatic image to obtain a single image that contains more information than any of the individual sources and the automatic color equalization algorithm is adopted by taking into account the color spatial distribution in the image.