Research On Image Super-Resolution Algorithms By Employing Image-Decomposition

The super-resolution image reconstruction technique is a hot research topic in the image processing research areas and is still an open problem. With the increase of the human demands for high resolution images, this technique has a lot of application areas, such as the computer vision, video surveillance, remote sensing, and medical images.The objective of the image super-resolution is to magnify the original image, enhance its resolution, and thus improved its quality. The super-resolution methods based on the training algorithms proposed recently have been considered as good methods and will have a bright future. These methods utilize the examples of the pairs of the high and low resolution patches as prior information to magnify an image.This thesis first introduces the theory and the development of the image super-resolution technique, and concentrates the research on the super-resolution algorithms based on examples. It proposes the super-resolution algorithm based on the image decomposition to improve the image quality after super-resolution. First, an image is decomposed into the cartoon and texture parts. Then, the image databases are built for the texture part and the cartoon part. The cartoon part is first magnified by methods based on examples and is then processed by using partial differential equation (PDE) methods, and the texture part is magnified by using the examples in the database. The final high resolution image is formed by the addition of the two magnified parts. This thesis is written by using the following structure.First, the database for the examples needs to be trained and acts as the prior information for the image reconstruction. This thesis proposes the method to individually form the example database for the cartoon and the texture parts. The cartoon part after the decomposition has mainly the low and middle frequency components of the image, while the texture part has mainly the middle and high frequency components. This not only makes the reconstruction image have more details and can improve the example matching accuracy during the reconstruction process.Then, it proposes the image super-resolution method based on the partial differential equations (PDE). This method can further process the magnified image to keep the image details while removing noises, and can achieve good quality.Finally, it proposes a super-resolution algorithm based on the edge enhancement. This method first checks the edge information of image patches and give large weights for these edge pixels during the image patch matching process and can improve the accuracy after image super-resolution. Simulation results have shown that the proposed method has better performance than the existing methods both visually and objectively, which are indicated by the SSIM, and PSNR measures.