Research On Multisensor And Multiresolution-based Image Fusion Algorithms

Image fusion, a special branch of data fusion, extracts the valid informationfrom multi-model images, to reduce the amount of data, to improve signal tonoise ratio with the use of the complementary information, to get morecomprehensive and more detailed scene information. Therefore, the study of howto achieve high quality images and get valid information from multi-modelimages is of vital significance. Before fusing images, there exits geometrictransformation between images to be matched, which need to be aligned. So theaccuracy of image registration is very important at the same time. The maincontent of this dissertation are as follows.(1) The study of preprocess in image registration, mainly consists of twoparts, the first part is to improve the quality of the image to be matched in orderto make it easy for feature extraction, and the second part is to extract contoursof objects in images as features. First, adjust the contrast of the image to makethe gray value distribution more concentrated, then use the color segmentationprinciple to get the target area, after that, considering to use a morphologicalopening operation to remove the discrete points noise outside the area and fill inthe empty area, finally exclude small areas by setting a threshold. With the use ofmorphological operation, we can extract regional contours as a collection iffeature points.(2) To formulate a solution for the image registration problem. The core ideais to model points as Gaussian mixtures, however, the model is built based on agiven set of feature points, so in order to build Gaussian mixture models, properfeature points must be extracted from the image. Here four feature point detectorare considered, namely, canny-corner operator, Harris-Laplace operator, SIFTand the contour shape operator. By optimizing the L2distance between models,the geometric transformation matrix can be estimated, and the experimentalresults are showed.(3) Research on the fusion algorithm between visible and infrared images.At first, image quality evaluation norm based on one or more images areintroduced, also five image fusion algorithm, which are respectively wavelet transform, spatial frequency, Laplace pyramid, PCA and the foreground-and-background way; then replace the wavelet transform with curvelets, which getbetter performance in describing curves, fuse the value channel with the infraredimage, then RGB image can be achieved with the inverse transform. Finally,apply six different fusion algorithms into two sets of images taken at differenttime and the fused images and assessment for them are showed.