GPU implementation for ℓ0-regularized blind motion deblurring

Blind image deblurring is a challenging problem in computer vision and image processing. Due to the highly computational complexity of the blind image deblurring, this paper presents an efficient parallel implementation that produces a deblurring result from a single image in a few seconds. The method is divided into a ℓ₀-regularized approach to estimate a blur kernel from the blurred image by regularizing the sparsity property of natural images and an improved TV-Deconvolution using split Bregman method to restore blurred image in the GPU section. The implementation makes effective use of commodity graphics processing units (GPUs). Specifically, we port the calculation of big vectors, matrices and FFTs to GPUs, perform intensive computations based on NVIDIA´s compute unified device architecture, and execute the rest of the operations related with control and small data calculations on the CPU. Experimental results demonstrate that blind image deblurring based on GPU runs an order of magnitude faster than the CPU, which is about 30 times the speed of the CPU, while the deblurring quality is comparable.