Research Of Super-resolution Reconstruction Technology On Remote Sensing

The spatial resolution is an important indicator for image quality. Higher resolutionbrings us higher amount of data, and also helps us get more useful information.However, digital image is corrupted by motion blur, undersampling, noise and otherfactors in the observation model. Multiframe super resolution imaging reconstructiontechniques supply us one way to conquer these problems and improve the imageresolution. It is one kind of technology using a series of low-resolution images capturedin the same sense to reconstruct a high-resolution image. Besides, sub-pixeldisplacement must exist among these low-resolution images. The technology has broadprospects in the military, medical, remote sensing, and many other fields.The super-resolution image reconstruction technique can be divided into twocategories: study-based image reconstruction techniques using single image andmultiframe super resolution reconstruction technology. This thesis studies the latter one.This thesis describes the super-resolution image reconstruction techniques basedon multiple remote sensing images. The range of application of Super resolutionreconstruction technology and the application particularity of remote sensing weredisscussed. This thesis entails solutions to two problems. One is the image registration.The other is the reconstruction of a HR image from multiple aligned LR images. Bothare important for the performance of super-resolution imaging. The thesis has severalmain contents:Registration algorithm includes frequency-based registration method, featurematching registration method, and optical flow estimation registration method. Byestablishing triangular network, feature-based registration method can be improved toachieve local deformation. While introducing the registration methods, experimentswere demonstrated to verify the actual effect of the various methods.Reconstruction algorithm: based on super-resolution reconstruction classificationsystem, multiframe super resolution is divided into frequency domain method,interpolation-based method and method based on probability theory (regularizationsuper resolution reconstruction). Taking projection onto convex sets as example, method can be applicated in local registration were analyzed. In the same way, experimentfollows every reconstruction algorithm introduction to verify the reconstruction result.Combining particularity of remote sensing image reconstruction, reconstructionexperiments results of ZY-3, resolution measurements chart of ISO and different remotesensing resources (TM and SPOT) using various were introducing in this thesis weredemonstrated and analyzed quantitatively. Finally, the influence of the number oflow-resolution images, spatial and temporal differences and the selection of thereference image on reconstruction effect was analyzed.