Research On Image Restoration Algorithm Based On Plug-and-Play Prior

With the advent of an information-based and intelligent society,in daily life and industrial production,digital images are becoming more and more important as the main channel for humans to perceive information.However,due to the complex imaging environment,the noise in the transmission of the equipment will often cause the degradation of the digital image.Although the microelectronics technology has made great progress in today's era,the pixels of imaging CMOS chips are getting higher and higher,and the noise interference is getting less and less.However,advances in hardware technology cannot completely solve the problem of degradation in the imaging process.Therefore,the main task of this article is to use computer software algorithms to solve the four common image restoration problems of denoising,deblurring,super-resolution and demosaicing.The algorithm in this article is implemented by python code.Among the existing methods,the method based on maximum posterior modeling has great influence.This type of method can model the process of solving the inverse problem of image restoration through maximum posterior modeling and Bayes theorem.The advantage is that compared with the deep learning method,the mathematical derivation process is clear,and the model is more interpretable.Moreover,the image degradation process is input as parameters into the model to participate in the solution,and a variety of degradation problems can be solved only by changing the degradation process parameters,and the model has a strong universality.The disadvantage is also very obvious: it needs to manually extract the prior features,and the effect is not as good as the effect obtained by the big data deep learning training.Moreover,the process of designing image block,similar block search and iterative derivation optimization in the solution process takes too long to process.The advantage of the method based on convolutional neural network is that the restoration effect is better,the forward reasoning process is faster,the principle is relatively simple,the technology is mature,and the complex reverse derivation training process has been integrated into the algorithm framework.The disadvantage is that the neural network belongs to end-to-end discriminative learning,that is,only one-to-one mapping can be obtained,so the model has poor universality and poor interpretability.Through the analysis of the solution based on the maximum posterior modeling,this paper proposes to insert the denoiser trained by the convolutional neural network as the prior information into the maximum posterior model.And complete the formula derivation of the variable separation algorithm to achieve this goal.Moreover,the method proposed in this paper combines the advantages of maximum posterior modeling and deep learning methods.The basic model established based on the maximum posterior makes up for the lack of universality of neural network methods,and the application of CNN denoiser makes up for the maximum posterior.The effect of the model method is poor and the processing speed is slow.Finally,the experimental results on the four image restoration tasks and the images collected by laser illumination prove that the image restoration effect and processing speed achieved by this algorithm are not only significantly better than the maximum posterior model method,but also compared with the method based on deep learning.Strong competitiveness.