Image Super-resolution Based On Deformable Patches

Vision is the primary approach for human perception,cognition and understanding from the surroundings.High resolution image is highly important for image demonstration,processing and analyzing.However,in the applications such as space surveillance,scenario monitor and military reconnaissance,the captured images are usually blurred with low resolution due to the device dimensions and costs.Meanwhile,in some cases,the capture is transient and unrepeatable,which means that the object of interest cannot be captured again by high resolution devices at any time.The existed image or video is the only source for image analyzing.Therefore exploiting the object information from the low resolution image or sequence and enhancing the resolution by super-resolution methods is a key scientific problem of great research value in computer vision and application.This work is supported by the National Defense 973 Project “** Coupling Mechanism Research”,State Key Program of National Natural Science Foundation “Space-Platform based Space Target Imaging and Recognition”and National High-Tech.(863)Project“Fast Processing and Motion Deblurring of **** Image”et al..And this dissertation focuses on the exemplar based image/video super-resolution.The key to this problem is the effective information utilizing for the ill-poseness in super-resolution.Specifically,this work investigates how to make the patch exemplars more expressive using patch deformable models.Moreover,the given high resolution image structures,sequence sub-pixel displacement information as well as the edge/texture priors are exploited for the image resolution enhancement.The main contributions and novelty of this thesis are as follows:1.For the problem of single image super-resolution using patch exemplars,a deformable patch model is developed in this work.Compared to the traditional methods,one patch is no longer a fixed vector but a flexible deformation field.Via the deformation of the patches,each patch can represent a bundle of patches with similar structures,which make it more expressive.As the experiments proved,the proposed model can produce the high resolution images more precisely using the dictionary of the same size.2.For the complex gradients patterns reconstruction problem,a deformable gradient decomposition method is proposed for single image super-resolution.In the view of the fact that structures such as single edges or corners are more invariant across different scales,the complex structures are decomposed into several such singular structures,each of which is deformable.Finally these patches are more suitable for high resolution estimation with more expressive deformable gradients dictionary.3.For the problem of fusing the edge structures and textures in single image superresolution,a method aiming at texture recovery combining with edge reconstruction is developed.First the different texture regions are segmented by local statistical features.For the edge structures,the deformable gradient composition method is applied.For each texture region,its type and scale are identified before introducing the high frequency texture details to the corresponding region.Finally the regions are merged by the gradient magnitude criterion.Experiments show the proposed method outperforms the convention methods under the structure texture similarity measures.4.For the problem of fusing sequence sub-pixel complementary information and the introduced external patch exemplars,the front-end and back-end deformable patches based methods are developed for image sequence super-resolution.In the front-end deformable superresolution,for the input patch in the current frame,the adaptive neighbourhood is constructed within which the dictionary patches are matched,deformed and weighted combined.In the back-end deformable patches based method,the deformable model is imposed on the sequence fusion result in order to introduce the external exemplar structures.These two methods are compared in the experiment on the suitable applying situations and the advantages to the convention methods are elaborated.