Spatial-invariant Fast And Robust Image Deblurring

Blurry images are universal in our life, there are many causes for the blur: the rapid moving of an object, shaking of the facility on other manual reasons such defocus. Natural image deblurring has been a hot research topic in recent years. In this thesis, a basic framework is proposed and most previous approaches tackling this problem can be unified under this framework.Firstly, a review of the most famous previous approaches is made in Chapter 2, including maximum a posterior method, ensemble learning method, regularization methods and edge enhancing methods.In following chapters, a unified framework is presented and it is shown that many previous methods can be understood as sharp edge enhancement with a nonlinear diffusion equation. A better viewpoint of the mechanisms and connections of previous approaches can be reached with nonlinear diffusion equations.For making better profiles for sharp images, shock filter is utilized to further enhance the diffused result. The strong edge profile is utilized for estimation of the point spread function(PSF)/blur kernel.For better speed and robustness, I propose 3 novel point spread function estimation algorithms, and a deconvolution method. PSF estimation algorithms are used to estimate the blur of the image, given the blurry image and sharp image profile. The deconvolution algorithm is used to recover the sharp image with the blurry image and the PSF.Thanks to the simple and robust algorithms plus scale-space optimization, the final framework can produce high quality results with little parameter tuning. This framework is interesting and useful both theoretically and practically. Future applications in many fields are promising.