Image Restoration Model Under Prior Constraint Regularization Framework

In the process of remote sensing imaging,affected by hardware equipment and external environment,remote sensing image will produce different degrees of degradation,which seriously affects the extraction and application of image information.In order to obtain high-quality clear images,upgrading imaging sensors to improve image quality has some problems such as high cost and complex imaging environment,so image restoration algorithm is usually used to improve image quality.The process of image restoration belongs to the solution of discomfort problem.Regularization method is widely used in the field of image restoration because it can improve the discomfort in image processing based on the prior information that can be introduced into the image.In this paper,the fuzzy and noise degradation problems in the process of image degradation are studied from two aspects of traditional methods and deep learning methods.Based on the regularization model framework,the traditional restoration method with multiple prior constraints and the interpretable deep learning restoration method based on L1 and L2norm are discussed.Specific contributions include:（1）The prior constraints reflecting the image characteristics can improve the image information corresponding to the prior in the image and stabilize the problem solution.At present,traditional image restoration algorithms only deal with the degradation of image noise or blur to improve image quality.Therefore,in order to better remove image blur,the group sparse constraint term is introduced,and the group sparse representation does not do special processing to the image noise.Therefore,the non-local mean self-similar constraint term is introduced,and a multi-prior constraint regularization model image restoration algorithm is proposed.Clear images are solved according to the prior knowledge of sparse and non-local mean self-similarity of image group.（2）For more prior constrained regularization model proposed in this paper in the process of its solution and regularization is hard to calculate model parameters（regularization coefficient,step length,etc.）need to be set,inspired by iterative threshold shrinkage algorithm,the regularization model combined with neural network,puts forward model of regularized image restoration algorithm based on depth of learning,A regularized image restoration model for optimizing L₁and L₂norms.The deep learning network solves the regularization model by replacing the regularization term with nonlinear transformation,and improves the interpretability of the deep learning network while optimizing the solution of model parameters.The multi-prior constraint regularization model image restoration algorithm proposed in this paper increases the PSNR value by 2.1737dB,2.0836dB and 2.1441dB on average for the images containing different types of blur and noise,indicating that the algorithm can remove image blur and suppress image noise at the same time.The regularization model image restoration algorithm based on deep learning was applied to different blur images,and the PSNR values of L₁and L₂norm algorithms increased by 7.9281dB,8.7005dB and 12.0028dB on average,which were all higher than those of L₁norm algorithm and L₂norm algorithm.Experiments show that the image restoration algorithm of regularization model based on deep learning is superior to the traditional multi-prior constraint regularization model,and the image restoration algorithm of regularization model with L₁and L₂norm is superior to the image restoration algorithm of regularization model with L₁and L₂norm alone.