Research On Blind Super-Resolution Algorithms For Image And Video

Improving the definition of images acquired from a certain digital imaging system has been a constant desire. Higher image definition usually means more details, which in one hand could improve the subjective quality and provide people with more information; in the other hand, they could improve the objective quality and make the following image segmentation and pattern recognition algorithms more accurate and robust. The most direct way to acquire high resolution images is to use high resolution sensors, but this would inevitably increase the system complexity and make it impractical for massive deployment. As a result, image super-resolution technique, which could improve the image definition by software based on existing hardware, has important values and significant meanings.Image super-resolution is a typical inverse problem. The establishment of a clear and accurate mathematical model for the forward path is usually a prerequisite for solving such inverse problems. In most existing super-resolution algorithms, it is often assumed that the image degradation model is known. However, it is not the case for practical application. The experiment results in this thesis have further proved that inaccurate image degradation model could badly affect the performance of most image super-resolution algorithms. Accordingly, the main topic of this thesis is the blind super-resolution problem with unknown degradation model.For single-image blind super-resolution, a novel iterative framework is proposed in which both the degradation model and super-resolved image are estimated alternately. Maximum-A-Posterior (MAP) estimation method is used to estimate the super-resolved image and example-based machine learning technique (KPCA) is incorporated to regularize the image degradation model. Intensive computer simulated experiments were carried out to validate the effectiveness of the proposed algorithm. The results show that the super-resolved images produced by the proposed algorithm are superior to traditional interpolation and non-blind super-resolution in both aesthetic and quantitative aspects.As the temporal extension of image signal, the video signal shares the same acquisition process, mathematical model and statistical character with the image signal. At the same time, it has its own unique feature: the strong temporal correlation. Based on such feature and the previously proposed single-image blind super-resolution algorithm, this thesis proposes a novel dynamic blind super-resolution algorithm for video sequences. The proposed algorithm exploits the temporal correlation of video signal and classifies the video frames as static frames and dynamic frames. For static frames, previously proposed single image blind super-resolution is directly used to produce the super-resolved image; for dynamic frames, image fusion technique with block classifier is introduced to combine the two high resolution candidates from spatial interpolation and motion compensation and produce high-quality super-resolved images. Experiment results show that the proposed video super-resolution algorithm could significantly reduce the computational complexity while maintain a relative acceptable performance. Meanwhile, preliminary analysis shows that the proposed algorithm also has considerable potentials for real-time application.