Image Deblurring Algorithm Under Non-uniform Noise Interference

Remote sensing images often contain blur degradation and non-uniform noise such as dead pixels and stripe noise.These two types of degradations are coupled to each other,damaging the image quality severely and limiting the application of remote sensing images.Therefore,the research of image deblurring under the interference of non-uniform noise is worthwhile and significant.Unlike a single non-uniform degradation or blur degradation,the coupling makes it difficult to remove this two types of degradations separately.This thesis analyzes and models the characteristics of stripe noise and dead pixels,as well as the blurry degradation,and establishes two noise/blurry correction frameworks to decouple non-uniform noise and blurry degradation,restoring clear images from degraded ones.Aiming to restore blurry images from interference of dead pixels,this thesis utilizes the outlier characteristic of dead pixels to construct a model which could accurately represent the nonlinear degradation process of dead pixels,and integrates the model into a joint optimization model for noise/blur correction,so that covers the shortage that the traditional dead pixels recovery methods rely on the position information of dead pixels.On the contrary,the position of dead pixels are adaptively estimated by the joint model,and then serve as a guidance of non-blind deblurring procedure.The result of synthetic dataset experiments and real-world dataset experiments represent that the joint correction method can effectively suppress the ringing artifact and improve the image restoration result by 2.32 d B of PSNR mean value compared with the independent dead pixel restoration and non-blind deblurring in sequence.For the problem of blind deblurring of images under the interference of stripe noise,this thesis adopts the strategy of combining data driving method and model driving method,utilizing the typical directional characteristics of stripe noise to establish a unidirectional variational regularization optimization model to estimate the blur kernel,and then correct the estimated blur kernel by a data-driven method.Finally,a joint noise/blur correction optimization model is constructed to perform the non-blind deblurring process under the interference of stripe noise.Experiments in this thesis show that the proposed deblurring algorithm is superior to other contrast algorithms by 2.79 d B of PSNR mean value under the interference of stripe noise.