Research On Blur Kernel Estimation Based On Region Selection

Image restoration technology has played an important role in many fields.Due to the limitations of various conditions,it is difficult to obtain specific degradation modes and blur kernels in the actual restoration process,and accurate estimation of blur kernel has a crucial role in the restoration of images.However,the proposed single-frame image restoration algorithms still have many deficiencies.Many of the methods are mainly focused on solving the image priors or observing the additional image deblurring,and little attention is paid to the influence of the image area information on the image deblurring.Many images have a large number of flat areas.This area not only lacks useful information,but also increases the cost of image restoration.In the non-flat areas of the image,there are also some areas that have not only no useful information,but also are not conducive to the accurate estimation of the blur kernel.In order to solve this problem and combine the needs of the projects,this paper proposes a new research on blur kernel estimation based on region selection.The main research content of this paper is to deblur the single image based on region selection,a new blur kernel estimation method based on region selection is proposed.The restoration of the space-invariant image is mainly through screening the zero-crossing points in the image,searching for feature points containing a large amount of useful information,calculating the gradient entropy in the calculated feature points to filter out a part of the interference points,and adding the Gabor filter to construct the feature vector.The feature vector finds the best area in the image under the CRF learning framework.In order to overcome the interference of noise on the blur kernel estimation,a weighted second-order difference regularization term is added to estimate the blur kernel.Some improvements are made to the maximum likelihood estimation recovery method.The nonlinear filter based on the maximum likelihood estimation is added to the image restoration to smooth the noise of the image and preserve the details in the image.A new spatially-variable model based on region selection is established for the empty variable image with slowly changing blur kernels.The empty variable image is divided into multiple sub-blocks for processing.The blur kernel estimation is performed on the gradient region for the best region of each sub-block.In order to enhance the edge of the image and reduce the ringing effect,the total variation regularization term is added to the restoration of the space-variable image,and it is solved by Split Bregman iteration.A series of simulated blur images and real blurred images experiments prove the effectiveness of the proposed algorithm.The results show that the blur kernel estimation method based on region selection proposed in this paper can efficiently estimate the blur kernel and obtain the restored image.The results show that the blur kernel estimation method based on region selection proposed in this paper can efficiently estimate the blur kernel and obtain high quality restoration images.For an empty variable image with slow blur kernel change,the space-variable restoration model based on region selection proposed in this paper is also applicable,and can obtain good recovery results.