Research On Image Motion Deblurring And Implementation Of GPU Acceleration

One of the long-standing challenges in digital imaging is motion image deblurring.Motion blur is generated from the relative motion between a scene and a camera duringexposure. While blur can be reduced by using a shorter exposure, this comes at anunavoidable trade-off with increased noise. Therefore, it is desirable to recover anun-blurred image computationally.Blind deconvolution is inherently an ill-posed problem due to the lack of constraint.To remove blur, we need to estimate the blur kernel and restore a natural looking imagethrough deconvolution. Blur kernel estimation is challenging because the algorithmneeds to distinguish the correct image-blur pair from incorrect ones that can alsoadequately explain the blurred image. In addition, optimization problems with millionsof variables in large scale deblurring problem should be solved in a very short time. Inthis work, we address a few aspects of these challenges.First, we analyze the blur estimation method that predicts the “sharp” version of ablurry input image and uses the two images together to solve for a blur kernel. We pointout the weakness of Extrema-based Multiscale Decomposition (EMD) algorithm used inthe prediction step. We further propose a novel method to overcome the limitation ofEMD algorithm, and show its benefits.Then, we exploit the ability of the modern graphics processors for solving imagedeblurring problems in a short time. We introduce NVIDIA’s Graphical Processing Unit(GPU) and the Compute Unified Device Architecture (CUDA) software platform. Wethen discuss some details of the GPU implementation. Computational results comparingthe speed of CPU implementation and the GPU implementation are presented that showthe advantage of GPU over CPU implementation.