Learning Hybrid Sparsity Prior For Image Restoration

Natural image is an important form of information carrier,obtaining high quality natural image has very important meaning in geological exploration,mapping,remote sensing,guidance and many other civil and military fields.Due to the influence of the limitation of hardware cost and environmental noise,current natural imaging technology is difficult to obtain satisfactory imaging results,the result images are often polluted by noise or blur,meanwhile,it's difficult to obtain high resolution image,it cannot meet to the requirement of high quality imaging results in practical application.Therefore,it is very meaningful to study natural image restoration(IR)algorithm and put it into practical application to improve image quality.Sparse coding based and deep learning based method for IR is relatively effective,but the former often requires multiple iterations,it takes a long time to solve the model,the latter has higher requirements on training data construction,and the generalization performance cannot be obtained for samples that do not appear in the training set.This paper dig natural image prior information based on the proposed Structured Analysis Sparse Coding(SASC)model,as we explored many types of natural image prior knowledge,we modeled solving the restored image from degraded image by construct several robust model regularizers.The main work and innovation points of this paper are described as follows.1.In order to make full use of prior information in external data,we improved the iteration based sparse coding model.Specifically,we approximate the solution of the restored image by a single gradient descend step,and then we expand the model iteration process including sparse coding step and restore image solving step as a neural network.On the one hand,the parameters in the model can be re-trained to achieve a better restoration performance,on the other hand,this training process also greatly reduces the computational complexity of our model.2.By mining natural image prior information,we proposed hybrid prior model which combines the structured prior information in degraded image and the prior information learned from massive data,we call our model as hybrid sparse prior model.The hybrid prior regularizer combines the model-based IR method and learning-based IR method,the structured prior is based on the fact that natural images has non-local self similar structured feature,and then we use deep convolutional neural network to learn the outside prior information from massive degrade-truth image pairs,a better image restoration performance is obtained by combining these two priors.3.Because our model has carried on modeling the degradation process of natural images,so our model can be applied to three IR sub-tasks including denoising,super-resolution,and deblurring,there is no need to design a network structure for a specific IR task,our model is a general IR model.We evaluate our proposed model on many public IR test dataset,the experimental results show that our method has outperform many existing methods on denoising,super-resolution and deblurring.