Blind Super-Resolution Reconstruction And Face Hallucination

Image resolution is not only a measure of the ability of an imaging system to distinguish the details of an image but also shows the fineness of details of the image. Since images with higher resolution can offer more details, there is a high demand for High Resolution (HR) mages. However, due to the limitations of the imaging system, images obtained in reality are usually degraded by blurring, noise and down-sample, thus with limited resolution. At present, there are mainly two ways to obtain HR images. One is to improve the performance of the hardware, while this may not be feasible for the high cost. The other way is to enhance the resolutions by signal processing technique, and one of them is the technique of Super-Resolution (SR) reconstruction. SR reconstruction is a technique which can obtain images with higher quality and resolution from a set of images with lower quality and resolution. It can overcome the inherent limitations of the low resolution imaging system and hence remove the affect of the blurring and noise. By far, SR reconstruction has been widely used in remote sensing, medical imaging, surveillance system and other areas.Generally, SR reconstruction technique can be divided into two categories: reconstruction-based SR reconstruction and learning-based SR reconstruction. As to the former, this dissertation mainly studies the blind super-resolution technique and face hallucination is discussed as to the latter. The dissertation is consisted of three parts as follows.First, the imaging process of the low resolution images is introduced and the corresponding mathematical model is also built. On this basis, the dissertation summarizes the classical algorithms of SR reconstruction technique, and focuses the attention on their basic principles, theory evidences, implementation procedures, the superiorities and deficiencies.Secondly, a blind super-resolution framework considering the sensor PSF is proposed. The sensor PSF is a kind of blurring caused by the fitness of a physical dimension in Low-Resolution (LR) sensors which is inevitable in the imaging process. In blind super-resolution, the negligence of the sensor PSF will make the identified blurring function disturbed by the sensor PSF, and hence decreased the accuracy of the blur identification and the quality of the restored HR image. Considering this, our scheme considers the sensor PSF and the identified blur separately, and builds the mathematical model of the sensor PSF which is accorded with the practical imaging process. And the model is implemented through the up and down-sample process. Then, Error-Parameter-Analysis algorithm is adapted to implement the blur identification and image restoration.Thirdly, face hallucination based on CSGT and PCA is proposed. Since face hallucination is a learning-based SR reconstruction technique, the hallucinated face images restored by this method have strong dependence on the training database. Therefore, in our scheme, all the face images (the input face image and the original training database) are transformed through Circularly Symmetrical Gabor Transform (CSGT) at first, and then local extremes criteria is utilized to extract the local features to represent these face images. Based on the local features, we calculate the Euclidean distances between the input face image and every face images in the original training database and the calculated distances are acted as the criteria of choosing the reasonable database. Once the training database is chosen, Principle Component Analysis (PCA) is employed to realize the hallucination.Finally, the problems to be solved related to SR reconstruction and future research topics are summarized. Furthermore, the prospect of the developing tendency is analyzed as well.