Single Image Super Resolution Using Discrete Cosine Transform Driven Regression Tree

Single image super resolution restoration is a process for solving the ill-posed problem of achieving a high resolution image from one low resolution image. This process is considered an active point of research due to the increased demands for high resolution imagery in many applications. This paper presents a proposed methodology for single image super resolution based on replacing the traditional discrete cosine transform basis. These bases are replaced by learned filters that can transfer the low resolution image from the spatial domain to high resolution coefficients in the discrete cosine domain. Accordingly, these estimated filters can then be applied to produce a high resolution image in the spatial domain through the standard inverse discrete cosine transform process. To learn these transformation filters two modifications in the decision tree algorithm are introduced to adapt the tree performance for the super resolution task. This is performed such that the node splitting decision depends on external features that minimize the regression errors. Experimental results demonstrate that the performance of the proposed algorithm is superior to that of the traditional interpolation single image super resolution method.