Research On Image Restoration Method Based On Deep Image Prior

Throughout the ages,image information plays a vital role in all aspects of people's life and work.However,due to the influence of various factors,the image quality cannot be guaranteed,which puts forward higher requirements for image restoration.Denoising and deblurring are the main imaging inverse problems in image restoration.At present,the more popular deep neural network methods usually use a large number of data sets to train the model to complete the image restoration work,but it cannot cope with the lack of data or the inability to obtain the original data.In addition,when the observation model for training does not match the model used for testing,the deep network method does not perform well.In response to the above problems,this thesis improves the Deep Image Prior(DIP)algorithm from the two perspectives of the data fidelity term and the prior term of the imaging inverse problem model,and further improve the performance of image restoration.First of all,the minimization model of the inverse imaging problem can be summarized as consisting of the fidelity term and the priori term.The DIP algorithm uses the deep neural network itself as a regular term.In this thesis,a Deep Image Prior algorithm combined with denoising regularization is proposed(DIP+RED).The U-shaped network structure of DIP is used to input a single degraded image and the restored image is obtained by training,which solves the problem of lack of training data sets.In addition,due to the ill-posedness of the imaging inverse problem,adding a regular term as a priori for degraded inversion will make the output result close to the original image.In order to make full use of the prior information of the input image,using the existing Non-Local Means(NLM)denoiser on the basis of the DIP method as regularization,and using the alternating direction multiplier(ADMM)to train the composite objective function of DIP+RED to avoid the explicit differentiation of the denoising function and produce a stable recovery.In order to ensure the efficiency of the algorithm,the network parameters are synchronized with the denoiser activation,and the denoiser is applied iteratively during the update process.Through comparative experiments,in terms of image denoising and deblurring and single image super-resolution,the improved DIP+RED method effectively improves the restoration effect of the original DIP algorithm,and the values of peak signal-to-noise ratio and structural similarity are increased.On the other hand,the fidelity term in the cost function of the inverse problem is used to make the output image conform to the observation model.The DIP method usually uses standard Least Squares(LS)as the fidelity term,which requires a lot of back-propagation iterations in the test.In this thesis,a Deep Image Prior algorithm using the fidelity term of Back Projection is proposed(DIP+BP).The use of the BP fidelity term can reduce the number of iterations and ensure that the projection of the optimized variable on the line space of the linear operator and the back projection of the linear operator applied to the observations are in agreement.In the optimization process,the stochastic gradient descent is used to minimize the objective function.The model does not require other external image data,which avoids the problem of mismatch between the training observation model and the test model.After experimental comparison and analysis,the improved DIP+BP method has a better restoration effect on image deblurring and image denoising and super-resolution than the original DIP scheme using LS as the fidelity item.