Super-resolution involves the use of signal processing techniques to estimate the high-resolution (HR) version of a scene from multiple low-resolution (LR) observations. It follows that the quality of the reconstructed HR image would depend on the quality of the LR observations. The latter depends on multiple factors like the image acquisition process, encoding or compression and transmission. However, not all images are equally affected by a given type of impairment. A proper choice of the LR observations for reconstruction, should yield a better estimate of the HR image, over a naive method using all images. We propose a simple, model-free approach to improve the performance of super-resolution systems based on the use of perceptual quality metrics. Our approach does not require, or even assume, a specific realization of the SR system. Instead, we select the image subset with high perceptual quality from the available set of LR images. Finally, we present the logical extension of our approach to select the perceptually significant regions in a given LR image, for use in SR reconstruction.