摘要: This paper presents a fast deblurring method that produces result from single image of moderate size in few seconds. We accelerate both latent estimation and kernel an iterative process by introducing novel prediction step working with derivatives rather than pixel values. In the step, we use simple processing techniques to predict strong edges estimated image, which will be solely used for estimation. With this approach, computationally efficient Gaussian prior becomes sufficient deconvolution estimate as small artifacts can suppressed prediction. For estimation, formulate optimization function using derivatives, numerical reducing number Fourier transforms needed conjugate gradient method. also show formulation results smaller condition system values, gives faster convergence. Experimental demonstrate our runs order magnitude previous work, while quality is comparable. GPU implementation facilitates further speed-up, making enough practical use.