From Holistic To Local Separating Strategy For Face Hallucination

Super-resolution reconstruction technology has broad application prospects andacademic research value both in the military, and in civil areas. And in the most fieldsof digital image application, image processing and analysis often require high resolutionimage or video as the ideal input signal. This makes super resolution technology plays aparticularly important role in preprocessing module of image application.In the thorough research of the super resolution algorithm of eigentransformationand the sparse representation-based super resolution algorithm, This paper aiming at thespecific problem of face super resolution (face hallucination) proposes our ownseparating strategy for face image super-resolution. Then on the basis of the newproposed method, this paper introduce a new choice mechanism of training library forenhancing the robustness of relative learning algorithm to the complex training libraryand the experiment results demonstrate the effectiveness of the new training setscreening mechanism. This paper mainly studies the content can be summarized asfollows:①This paper first introduced the academic research significance and applicationvalue of super resolution problem research, analyzes the Present Situation ofsuper-resolution reconstruction technique, especially the super resolution technology offace image category, domestic and abroad. And then the paper summarize the threecommon type of super-resolution reconstruction technology and their representativealgorithms as well as the special algorithms for face hallucination problems, analyzedthe difficulty of research subject and lay theoretical foundation for the subsequentresearch.②This paper more deeply analyzes the modeling of image degradation from maththeory and the key techniques of super resolution. The paper introduces why thedegradation of image quality process will happen, and use proper mathematical methodto descript the process. And then the paper introduced the more general unifiedexpression of the image degradation model in mathematics currently, try to understandthe theoretical background of problem and find the suitable methods to solve theproblems. In addition, we analyze several current commonly used objective imagequality assessment standards from the mathematical perspectives. ③Targeting the specific face hallucination problem, this paper proposes a novelseparating strategy for face hallucination in order to better introduce the priorsinformation of the face image category for improving the outgrowth face image quality.This strategy is aimed at the face image hallucination given one single low resolutionface input image. First this paper puts forward a local patch based eigentransformmethod. It can better introduce the prior information of face image category into therecovery of the whole facial structure and apply the high and low resolution face imagetraining set pair to zoom the input low resolution face image to a medium resolution.Then we use the patch-based sparse representation algorithm and the relearning highand low resolution overcomplete dictionary pair to reconstruct the fine details of themedium resolution face image. Finally, the related experiments and output image hasshown the superiority of the strategy④In the learning type of super resolution reconstruction methods, How extractthe feature information to better find the structure and details similar to the lowresolution input face image in the training set is also a key problem. This paperelaborates the first-order gradient and second-order gradient operators in the learningsuper resolution methods, their original form and their derivation process, Then,according to the multi direction characteristics of the atoms of the sparse representation,the paper ameliorates the feature extraction process in the learning tpye facehallucination method, add the four gradient feature extraction operators to eight onesand propose a new feature extraction method. And it can better achieve the similarinformation of facial structure and details. The experiment proves the effective of ourmethod.⑤Since the learning type super resolution method use the face holistic structuresand local details as the main pirors, the face image samples from different categoriessuch as different races play an important role in the effectiveness of super resolution.Therefore with the respect to face from different races, we proposes a new training setscreening mechanism based on2norm distance along with the proposed separatingstrategy for face hallucination to solve the problem that the training set composition willseverely affect the performance of the learning type super resolution methods. Finallythe experiment prove that introduction of this mechanism to the learning type superresolution method will enhance the robustness of the method to the training set.