On image-driven choice of wavelet basis for super resolution

We move closer to deriving an image-driven criteria for the choice of wavelets for single image super resolution (SR). We start with the hypothesis that higher edge densities are better reconstructed by wavelets with higher number of vanishing moments and smaller support size. We examine SR performance of different wavelets on image categories with different amount of details. We use the slope of log power spectral density as a proxy for edge density. The results strengthen our hypothesis. We also interpret the results of two metrics and conclude that SSIM is a better perceptual metric than PSNR because it measures preservation of edge reconstruction. We use a TCEM model for estimating detail coefficients due to its previously reported superior performance.