Research On Image Restoration Technology Under Incomplete Sensing

During the acquisition,transmission and reception of image information,partial pixel loss is often caused by uncontrollable factors such as masking and selective transmission,which can affect the final visual experience.This phenomenon is known as incomplete sensing.Therefore,image restoration techniques under incomplete sensing have significant practical value.Image information can be represented by matrices.Actual image matrices often exhibit low-rank characteristics,and thus the image recovery problem can be approximately regarded as a low-rank matrix approximation problem,i.e.,solving the rank minimization problem.Since rank minimization is an NP-hard problem and difficult to solve,the convex substitution of rank functions represented by nuclear norms is widely used.In recent years,the truncated nuclear norm,as an improved non-convex function of nuclear norm,has been considered to better approximate rank functions.In Chapter 3of this thesis,we propose two methods for low-rank matrix completion with truncated nuclear norm minimization to the image recovery problem under incomplete sensing.First,we propose a truncated nuclear norm minimization method with extrapolated proximal gradient.In order to further reduce the number of iterations in solving the constrained convex optimization problem in truncated nuclear norm minimization,we adopt a momentum-based gradient descent method to solve the convex optimization problem.The gradient descent method has low computational complexity,and the introduction of momentum mechanism can effectively suppress oscillation and enhance the stability of the algorithm.The experimental results show that the proposed truncated nuclear norm minimization method with extrapolated proximal gradient performs well in the image recovery quality under incomplete sensing.The proposed truncated nuclear norm minimization method with momentum gradient descent performs well in terms of image recovery quality,iteration time,and robustness of truncated singular value numbers.During image and video transmission,pixel block loss often occurs when decoding compressed image and video streams,which is a typical image loss problem under incomplete sensing.The error concealment is an important method to solve this problem.Among them,spatial error concealment only uses the relationship between blocks in the current frame image for recovery,without using information from other frames,so it has significant research value.Among many spatial error concealment algorithms,those sparsity-based algorithms utilize the sparsity of the image and perform better in cases of drastic changes of image content when compared to interpolation-based methods.Therefore,Chapter 4 of this thesis focuses on researching sparsity-based spatial error concealment algorithms and optimizing the local region matching stage and the local linear correlation model establishment stage.In the local region matching stage,we propose an adaptive threshold-based method that can flexibly adapt to lost subregions with different features,thus providing more accurate observation space and potential space for dictionary construction and local linear correlation model construction.In the local linear correlation model establishment stage,we propose a kernel ridge regression method based on the α-ML kernel function(Laplacian kernel norm under Minkowski distance)as the local linear correlation model.Compared with the Gaussian kernel function,the α-ML kernel function has lower parameter sensitivity and better flexibility.Experimental results show that the proposed algorithm outperforms existing spatial error concealment algorithms in image recovery quality under various typical block loss scenarios.When conducting research on image restoration under incomplete sensing,it is necessary to perform experiments on a large number of images and compare the performance of various image restoration algorithms.We have developed a visual image restoration system for the problem of image restoration under incomplete sensing,aimed at reducing the workload of researchers.The system integrates the performance analysis functions for both single-algorithm image restoration and multiple-algorithm image restoration,which is designed with different modules for different application scenarios,and allows for efficient selection of the optimal image restoration algorithm for the current scenario.