אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Super resolution (SR) algorithms typically assume that the point spread function (PSF) of the camera is known, or else assumed to be a standard low-pass filter (e.g. a Gaussian). We demonstrate that the performance of such methods significantly deteriorates when the PSF deviates from their “one-size-fits-all” model. Deviations from the nominal PSF may be caused, e.g. by small camera shake inducing motion blur, by optical variations between sensing devices, etc. Such variations may be hardly distinguishable in the low-resolution image, but may strongly affect the quality of the high-resolution output. In this work, we propose a general framework for “blind” super resolution. Our approach can estimate the blur kernel either with the aid of an external database of high-resolution example images, or by relying directly on the internal statistics of the single low-resolution image. In particular, we show that recurrence of small patches across scales of the low-res image (which also forms the basis for single-image SR), can be used for estimating the optimal blur kernel. We show that plugging our estimated kernel into existing super-resolution algorithms results in improved reconstructions that are comparable to using the ground-truth kernel.

Joint work with Michal Irani.