Solving Inverse Computational Imaging Using Deep Generative Gradients Of Priors

Reconstructing original signal from the compressed sensing signal has always been a hot issue in the field of computer imaging,the core of which is how to improve the imaging quality.Compressed sensing signal reconstruction is an ill-posed inverse problem.There will be various challenges in the reconstruction process,and the most common ones are edge blur and noise interference.Recently,deep learning methods have shown great potential in solving various inverse imaging problems.However,due to the particularity of different imaging systems of computational imaging,it is usually necessary to model different imaging inverse problems separately,which lacks versatility and flexibility.To this end,this paper mainly studies computational imaging methods based on unsupervised deep learning,using statistical prior distribution of the generated network learning image,and then integrating it into the iterative reconstruction process.The extracted explicit prior has good versatility and can be used to solve a variety of imaging inverse problems.The generative model based on scoring network learns the gradient of the data distribution from training set,and then the Langevin dynamic annealing will be used to update the Generative Gradient of Priors(GGP)of the trained generative model.This is the first time that a prior network based on score matching network and Langevin annealing method has been introduced into the field of computational imaging.The generative model proposed in this paper doesn't estimate data distribution,but the data distribution.In the imaging reconstruction stage,the annealed Langevin dynamics is used to iteratively obtain the estimated value of the previous gradient.At the same time,each iteration will update the data consistency constraints.This method can show good adaptability in multiple imaging systems.Compared with the existing methods,the gradient prior method of the generative model proves its applicability in the three computational imaging problems.What's more,compared with traditional algorithms and existing deep learning methods,the result of the gradient prior of the generative model shows a better reconstruction effect.