No-Reference Noise and Blur Detection via the Fourier Transform

This research presents a new method of distinguishing between noisy, blurred and otherwise uncorrupted images via the Fourier transform. The spectrum of an image corrupted by noise is markedly different from one that is marred by blurring, or one that is not damaged at all. In particular, there are more high frequency terms in the spectrum of a noisy image than in that of a blurred image, and, to a lesser degree, that of an image that has not been altered. So as to better emphasize these distinctions, a cumulative distribution function of the terms of the spectrum of an image is assembled. Using only this construct, a new technique that enables a machine to reliably assess the number of high frequency terms in an image, and therefore differentiate among the three types of images, is brought forward.