Research On Super Resolution Reconstruction Algorithms

The purpose of super resolution(SR) is to reconstruct a high-resolution (HR) image with a higher quality version effect from one or a sequence of low-resolution (LR) images, which are captured from the same scene and sub-sampled as well as shifted with sub-pixel precision. The resolution enhancement technique has been recently utilized widely, for example, the computer version, remote sensing, safety area and image processing. With the research of image degradation model, many SR reconstruction methods in the spatial domain become more and more popular. However, these methods do not pay enough attention to the situation where the size of target in the zooming image sequence changes greatly.In this thesis, we focus on super-resolution reconstruction methods of zooming image sequence and rotated image sequence. Firstly, based on the LR image sequence, zooming degradation model are present. A target-size image produced from LR image by interpolation is then selected as the initial HR image. Based on the degradation model and the matching points generated by scale-invariant feature transform (SIFT) respectively in HR and LR image, the transform formula is generated. By this transform formula, an expected LR image is generated from the HR image. Compute the difference between the simulated LR image and the real observed LR image mentioned in computing transform formula. When the difference is larger than some threshold, the difference is added into the HR image by back projection so that the HR image is updated. If the iterative time reaches the threshold, the whole process is ended. Otherwise, the some operations are employed on other LR images.As to rotated image sequence, a LR image was chosen and enlarged to the target size, which is taken as initial HR image. Compute the feature matching points between HR image and another LR image and then rotate the LR image by the rotation matrix. By degradation model, compute the expected LR image from the HR image and the difference image between expected LR image and real LR image is produced. When the difference is larger than some threshold, the difference is added into the HR image by back projection so that the HR image is updated. If the iterative time reaches the threshold, the whole process is ended. Otherwise, the some operations are employed on other LR images.Peak signal to noise ratio (PSNR), structural similarity index (SSIM) and energy ratio between high-frequency components and the whole image are employed to evaluate the image quality in this thesis. PSNR and SSIM represent the similarity between two images while energy ratio between high-frequency components and the whole image means the proportion of image details.Finally, the improved algorithm is employed on simulated zooming LR image sequence, real zooming LR image sequence, simulated rotated LR image sequence and real spun LR image sequence. PSNR, SSIM and energy ratio between high-frequency components and the whole image are calculated to judge the quality of our rebuilt HR image. The experimental results show that the HR image reconstructed by our proposed SR algorithm with zooming LR image sequence or rotated LR image sequence does have a good quality.