Research On Point Spread Function Estimation And Super-resolution Reconstruction Of Defocused Image

Image is an important medium for human being to obtain external information.High resolution(HR)image is helpful for people to perceive and understand things outside.However,due to the limitation of imaging device detector and the influence of the external environment,image degradation often happens.How to improve the resolution of images based on existing imaging devices is a hot topic in the field of image processing.Image super-resolution reconstruction is an effective way to solve this problem,but existing super-resolution reconstruction algorithms only focus on the well-focused images ignoring the value of the defocused images in the focusing process.In fact,these blurred defocused images are also observations of the scenes,and also contain valuable target information.This paper aims to mine the observation information contained in the defocused blurred images to realize image super-resolution reconstruction and the main work of this paper is as follows:(1)The method of point spread function(PSF)estimation for defocused image is studied,and a PSF estimation method based on particle swarm optimization(PSO)algorithm is proposed.Because only the low-resolution(LR)blurred image is known,the PSF corresponding to the HR image cannot be estimated directly.Therefore,from the perspective of the image blur,the PSF relationship between LR and HR images is discussed and the equivalent transformation relationship between the two is established;Then,based on the image blur process,the cost function of PSF estimation is established and according to the disk model of PSF,the estimation problem of PSF is converted to the optimization problem for the blur radius and ultimately achieve the estimation by the PSO algorithm.Experimental results show that the proposed algorithm can accurately estimate the PSF of defocused images.(2)Combining the estimation of PSF with the reconstruction of the image,a PSF estimation method based on alternate iteration is proposed.In the PSF estimation section,the HR image remains unchanged;in the image reconstruction section,the PSF remains unchanged;then update the PSF and the HR image by alternate iterations until the algorithm converges.The experimental results show that the algorithm can accurately estimate the PSF of the defocused image,and choosing the smooth region of the image to estimate the PSF can obtain faster convergence speed and higher estimation accuracy.(3)In terms of dictionary learning,the traditional method is generally to select a natural image library that has nothing to do with the reconstructed image as a dictionary learning sample,and then learn a global dictionary with a wide range of expression capabilities.It is not only time-consuming,but also inefficient in dictionary learning,and the obtained global dictionary is less adaptive.In order to reduce the dictionary learning time and improve the learning efficiency,the paper constructs a pyramidal sample training set.Each layer of the pyramid is interpolated from the image to be reconstructed,which not only greatly reduces the number of samples but also integrates the multi-scale structural self-similarity of the image into the dictionary learning,thus improving the ability of the dictionary to express the reconstructed image.Besides,in order to improve the adaptability of the dictionary,the paper adopts the way of classification dictionary learning.Firstly,the K-means classification method is used to classify the training samples;then the principal component analysis method is used to learn the corresponding sub-category dictionary for each type of sample;in this way,multiple types of dictionaries can be learnt.Then in the reconstruction process,the most matching sub-category dictionaries can be selected adaptively according to the different image structure types.Therefore,compared with a single global dictionary,the adaptive dictionary can express the structural information of the image more accurately.(4)Aiming at the ill-posedness of image super-resolution reconstruction problem,the nonlocal means regularization constraint and the bilateral total variational regularization constraint are introduced,which improves the image reconstruction quality a lot.According to the phenomenon that multi-scale structural similarity of images is prevalent in natural images,it is used as an a priori constraint for image super-resolution reconstruction,and a nonlocal means constraint regularization term is constructed.In addition,the bilateral total variation is also introduced which can constraint the image from both the spatial relationship and the gray value relationship of the pixel,and can well maintain the edge information of the image.Experimental results show that the proposed algorithm achieves good results in image super-resolution reconstruction under the above dual regularization constraint framework.