Image Restoration Via Universal Denoising Networks

In nowadays information society,images are not only indispensable media for human,but also widely used in all walks of life,including medical diagnosis,aerospace and so on.It will greatly affect its value if image quality degrades for a variety of reasons.Therefore,image restoration technology is critical and needs to be developed.In fact,image restoration is an ill-posed problem from a mathematical point of view.There are many image restoration methods proposed so far.In this thesis,we mainly consider two categories: building complicated optimization models for image prior and learning image prior using discriminative methods,with different advantages and disadvantages.Those methods based on model optimization take the advantage of the flexibility in solving multiple problems.However,they usually suffer from many iterations and amount of computation as well as long wall-clock time.Different from those methods based on model optimization,the discriminative methods,especially those use deep neural networks with end-to-end training,can get better restoration results and more importantly,they shed some light on real-time image restoration.Whereas,these methods are specialized for certain purpose because of the single-task training.Therefore,the research searching a trade-off which shares advantages of both methods make sense.In this thesis,we build the model for image restoration problem using Alternating Direction Method of Multipliers(ADMM)that disentangles the image degradation term and image prior term.On one hand,the image degradation is not involved in the step of learning image prior essentially an image denoising problem,we use convolutional neural networks to solve image denoising problem,thus our method is very flexible.On the other hand,the performance of convolutional neural network for image denoising is critical to the whole image restoration problem.We introduce a global denoising neural network based on mathematical theory to handle image denoising problem,by training gray and color images denoisers with different noise levels.The denoisers will be plugged in the denoising sub-problem of the ADMM algorithm to solve image restoration problem.In this thesis,we focus on three classical image restoration problems,including image denoising,deblurring and single image super-resolution reconstruction.Numerical experiment results show that the proposed method not only performs competitive with based on deep learning methods,breaks the limitation that can only deal with a single problem,but also the computation time is much less than those model-based optimization methods.