Image Restoration Algorithm Based On Deep Learning

Image restoration is one of the typical questions in computer vision processing.For example,in the field of natural images,the imaging process is often affected by gaussian noise,salt and pepper noise,etc.,and the image transmission is often accompanied by information loss and resolution reduction.In the field of medical image,magnetic resonance imaging,for example,although the high-contrast clear images can be obtained without ionizing radiation or contrast agent in the imaging process,in order to minimize the acquisition time and improve the efficiency of diagnosis,scholars have been working on the rapid reconstruction of magnetic resonance,and the undersampling will often produce artifacts and affect the final diagnosis.The fundamental goal of image restoration is to preserve the details of the image while removing noise and other interferences.In recent years,the rise of deep learning has provided a new idea for image restoration.Based on the image restoration algorithm and the construction of deep learning network,this research is carried out on the restoration of deep learning network on natural images and medical images respectively:(1)For natural image denoising with pepper and salt under supervised deep learning.To solve the problem of salt and pepper noise denoising in image acquisition and imaging,we propose a denoising algorithm based on supervised deep learning network.The classical convolution model of dense connection is used to construct the network,the function of convolution layer as distribution transformation and pre-activation batch normalization is discussed.All the experimental results show that the proposed network algorithm has better image restoration performance than the comparison algorithm.(2)For rapid Magnetic Resonance image reconstruction based on the prior-driven information of unsupervised deep learning.To speed up MRI reconstruction by reducing data acquisition,we propose two fast parallel MRI reconstruction methods driven by unsupervised deep learning prior information.Based on the experience of supervised learning network construction,a denoising auto-encoder model is trained to extract the prior information,and the extracted prior information is applied to different iterative networks to achieve the goal of rapid reconstruction.Deep learning network is used to train the learning ability of prior information,which is combined with the parallel imaging iterative model based on compressed sensingThe experimental results show that the proposed network has good performance.In the natural image denoising task,the image with clear texture can be obtained even when the noise density is very high.In the task of rapid magnetic resonance reconstruction,we can see that the proposed algorithm retains clearer texture even when the sampling rate is low,and the optimal result is obtained when the objective index is used for quantitative evaluation,which achieves rapid and high-quality reconstruction and has certain clinical and engineering application value.