Deep Learning Based Sparse Coding For Image Restoration

With the coming of big data and artificial intelligence era,the demand of clear image becomes more and more important in field such as surveillance,remote sensing,self-driving cars and so on.However,there is a sharp increase in the volume of image and video data but,meanwhile,the quality is not good enough.All these things mentioned above have brought a great challenge to image data transmission and processing missions.At present,the traditional methods of image restoration are calculated with the phenomenon of high complexity.In order to solve this problem,we use deep learning based sparse coding model to speed up the image restoration algorithm.The model has high efficiency in solving the sparse linear inverse problem.More concretely,the model used in this paper “unfolds” the soft threshold based approximate message passing(AMP)algorithm and form a feed-forward neural network.The structure of the resulting learned AMP(LAMP)network is similar to that of learned iterative shrinkage thresholding(learned ISTA,LISTA)algorithm but contains additional “bypass” paths whose gains are set in a particular way.The bypass paths in LAMP algorithm have a different topology and a different gain-control mechanism.We show numerically that LAMP’s outputs are more accurate than those of LISTA.Based on these characteristics,we propose an image reconstruction algorithm under compressed sensing and an image super resolution algorithm based on the deep learning based sparse coding model.The algorithms are as follows:Firstly,we apply the deep learning based sparse coding model to the image reconstruction problem under compressed sensing and propose a natural image reconstruction algorithm.The reconstruction algorithm is a supervised learning algorithm and a sparse encoder is optimized through a large number of training samples.After training,the encoder can be used to predict the approximate sparse codes of image transform domain with a fixed computational cost and a prescribed expected error.The predicted sparse codes will then be used to obtain the final reconstruction image by inverse transforming.The trained encoder can be used for image reconstruction under compressed sensing of corresponding observation rate.The iteration number of the trained encoder is less than 10,so this is a rapid image reconstruction algorithm under compressed sensing and has a wide application prospect and popularization value.Secondly,we propose a super-resolution algorithm based on deep learning based sparse coding model.Inspiring by the idea of recurrent neural networks,we propose an end to end recursive image super-resolution network.First of all,with the application of AMP algorithm,we have implemented an image super-resolution feed-forward neural network whose steps strictly corresponding to the traditional sparse coding method.And the existing sparse representation then can be effectively combined in our network structure.Then,we introduce the idea of recursive neural network and propose a multi-level model for image super-resolution.Also,we use transfer learning as our training method to improve model training efficiency.In this paper,we propose an image reconstruction algorithm under compressed sensing and an image super resolution algorithm based on deep learning based sparse coding model.The experiments running on test datasets show that our reconstruction algorithm overcomes the defect of slow reconstruction speed and achieves better recovery results.