Comparison of single and multi frame super resolution reconstruction algorithms with analytical interpolation methods

Abstract

Generating high resolution image (or image sequence) from low resolution image (or image sequence) has various applications such as image expansion, printing and conversion between different resolution formats. There is a huge amount of study on HR image reconstruction problem in the literature. These methods can be broadly divided into two main classes: analytical reconstruction techniques and super resolution reconstruction techniques. In the former case, reconstruction is established using an interpolation kernel. The original LR image is convolved with interpolation kernel to obtain the continuous data. Then continuous signal is sampled again according to the desired resolution. On the other hand, in SR reconstruction, the idea is to fuse different samples obtained at different time instants from the same object by a single camera. SR reconstruction can be posed in another way. HR image can be reconstructed by combining different samples obtained at the same time instant from the same object by multiple cameras. There should be sub-pixel shifts between sampling locations to make the reconstruction possible. In this M.S. Thesis, subjective and objective comparison of 5 different analytical interpolation methods and 2 super resolution image reconstruction methods is given. An adaptive filtering approach least mean squares (LMS) filtering and robust super resolution are used for super resolution image reconstruction. SR methods are compared with bicubic interpolation, wavelet based interpolation, edge adaptive interpolation, interpolation using wide sense Markov random fields and interpolation using exponential based kernels. All seven methods are tested on different videos and frames. PSNR and SSIM measurements are given. Also, subjective tests are conducted on the experimental results.