Real-time image deconvolution on the GPU

Two-dimensional image deconvolution is an important and well-studied problem with applications to image deblurring and restoration. Most of the best deconvolution algorithms use natural image statistics that act as priors to regularize the problem. Recently, Krishnan and Fergus provide a fast deconvolution algorithm that yields results comparable to the current state of the art. They use a hyper-Laplacian image prior to regularize the problem. The resulting optimization problem is solved using alternating minimization in conjunction with a half-quadratic penalty function. In this paper, we provide an efficient CUDA implementation of their algorithm on the GPU. Our implementation leverages many wellknown CUDA optimization techniques, as well as several others that have a significant impact on this particular algorithm. We discuss each of these, as well as make a few observations regarding the CUFFT library. Our experiments were run on an Nvidia GeForce GTX 260. For a single channel image of size 710 x 470, we obtain over 40 fps, while on a larger image of size 1900 x 1266, we get almost 6 fps (without counting disk I/O). In addition to linear performance, we believe ours is the first implementation to perform deconvolutions at video rates. Our running times also demonstrate that our GPU implementation is over 27 times faster than the original CPU implementation.