Fast Two-Step Blind Optical Aberration Correction

AbstractThe optics of any camera degrades the sharpness photographs, which is a key visual quality criterion. This degradation characterized by point-spread function (PSF), depends on wavelengths light and variable across imaging field. In this paper, we propose two-step scheme to correct optical aberrations in single raw or JPEG image,i.e., without prior information lens. First, estimate local Gaussian blur kernels for overlapping patches sharpen them with non-blind deblurring technique. Based measurements PSFs dozens lenses, these are modeled as RGB Gaussians defined seven parameters. Second, remove remaining lateral chromatic (not contemplated first step) convolutional neural network, trained minimize red/green blue/green residual images. Experiments both synthetic real images show that combination two stages yields fast state-of-the-art blind aberration compensation technique competes commercial algorithms.KeywordsPoint-spread functionOptical aberrationsBlind deblurringSpatial filterEdge non-linear filtering