Single‐image super‐resolution using online kernel adaptive filters

Abstract

The online kernel adaptive filters are non-linear filters which provide impulse response and are more efficient compared to other kernel algorithms. The performance of kernel adaptive filters depends on dictionary size. Here the single-image super-resolution using online kernel adaptive filters is a learning-based method. The algorithm generates a sparser solution for obtaining high-resolution image from a low-resolution image. It finds out a dictionary with most significant set of basis vectors using the spatial similarity among the dictionaries created from the low-resolution and high-resolution image patches in the training set. The dictionary is utilised to generate the high-resolution image. The algorithm is analysed on three different kernel adaptive filters, extended kernel recursive least squares, kernel recursive least squares tracker and naive online regularised risk minimisation algorithm. The performance of the super-resolution method is evaluated on a large number of images and is compared with the state-of-the art non-linear solutions to the super-resolution. The results show a better progress in peak signal-to-noise ratio up to 1.2 dB.