Multi-noise And Multi-channel Derived Prior For Image Restoration And MRI Image Reconstruction

Image restoration and reconstruction is a basic and extensive research field,aiming to eliminate or reduce the degradation of image quality.As an under determined inverse problem,although many excellent algorithms have been developed on how to mine prior information,most of them still have some limitations,such as only suitable for a single custom restoration task,long iterative reconstruction time or unstable results.With the development of deep learning,convolutional neural network has been widely used in image restoration and reconstruction because of its powerful ability of feature learning and expression.In this paper,under the guidance of the core theory of convolutional neural network and image prior information,we study different image restoration tasks,such as image deblur,compressed sensing reconstruction,MRI image reconstruction,etc.,and propose a new general algorithm to improve the limitations of most algorithms,so that the training model can be applied to multi task scenes,such as in image de blur task.The model is suitable for different levels of blurred image restoration,and for different sampling modes and sampling rates in compressed sensing reconstruction and MRI reconstruction.The main research contents include:(1)Through the implementation of multi noise and multi-channel network learning to expand DMSP,an enhanced and more robust network guidance iteration method is proposed.Specifically,by learning a DMSP deep learning network with multi-channel noise model,valuable high-dimensional prior knowledge is extracted from the image with multi-channel,and then the high-dimensional network prior information is integrated into the iterative reconstruction process by using the variable enhancement technology.Next neighbor gradient method and alternating iterative optimization method are used to deal with the reconstruction process.Through a large number of experiments,the advantages of the algorithm in image de blur and compression perception reconstruction are verified.(2)MRI reconstruction based on a priori information learning model guided by multi-channel and multi noise network is realized.On the basis of preserving the model applied to image de blur and compressed sensing reconstruction,two sets of different medical image training data sets,brain image data set and knee image data set,are used in the training process.Through the strong representation ability of the model itself,it can achieve more accurate restoration effect on the data set with less training data and easier to obtain.The multi-channel and multi model prior information network proposed in this paper has been evaluated in different image restoration tasks,and compared with the latest methods.The experimental results show that PSNR,SSIM and HFEN are significantly improved in the application scenarios of image de blur,accurate compressed sensing reconstruction and MRI reconstruction.Even under the condition of high blur core and high undersampling rate,it can provide convincing clear image.To sum up,based on the multi-channel and multi-model prior information network,this paper uses the variable enhancement technology to integrate the high-dimensional network prior information into the iterative reconstruction process as the main core point,and uses the proximal gradient method and the alternative iterative optimization method to deal with the MEDMSPRC algorithm of the reconstruction process,which effectively overcomes the shortcomings of the existing methods.In image de blur,compressed sensing image reconstruction and MRI,the reconstruction has achieved good results,which opens a new way to explore the application of high-dimensional network induced prior information in image restoration and reconstruction.