UTH denoising benchmark

This dataset consists of approximately 43.5K images of various levels of complexity, corrupted by noise of various levels of 5 data-independent noise types, i.e. Salt and Pepper, Gaussian, Laplacian, Uniform and Speckle. Overall, the generated dataset consists of pure noise images, i.e. homogeneous images of 128 intensity, corrupted by the 5 noise types, as well as of 6 sets of images, which are generated by introducing each noise type to 6 real-world images (namely 'saturn' and 'eight', which are low complexity images, 'gantrycrane' and 'I20', which are medium complexity images and 'greens' and 'I14', which are high complexity images). The sampling of the noise parameters is performed such that successive samples are very close to one another, for the case of pure noise images, so that in the entropy\u2013complexity plane the corresponding graphs are pseudo-continuous. The remaining images in the dataset are obtained by corrupting each of the 6 real-world images with the same noise samples from each of the 5 noise types that were used to generate the pure noise images.