Research Of Image Restoration Methods Based On Prior Information Of Convolutional Neural Networks

The quality of digital images is not only related to image visual effects,but also plays a key role in many fields such as medicine,astronomy,surveillance and detection.The research of image restoration technology aims to eliminate or reduce image quality degradation,and is one of the research hotspots in the fields of digital image processing and computer vision.As a key factor in dealing with ill-conditioned problems in image restoration,image prior information helps restore high-quality digital images and plays a vital role in various computer visual tasks.With the deepening of research in the field of deep learning,convolutional neural networks have attracted extensive attention at home and abroad with their powerful feature learning and feature expression capabilities,and have been effectively applied in the research of image restoration technology.In this thesis,we studied different image restoration tasks such as image deblurring and image super-resolution reconstruction under the guidance of the core theory of convolutional neural networks and image prior information.The research results mainly include the following aspects:(1)An image deblurring algorithm based on convolutional neural networks is proposed.On the basis of the residual network,we further increase the number of network channels,learn the distribution features from the noise observation to obtain the image prior information,and insert the learned network models into the iterative optimization framework to solve the image to achieve the image deblurring.In the proposed algorithm,we use the half-quadratic splitting method to solve the problem.The experimental results show that the proposed algorithm achieves better image deblurring effect,but there is still instability.(2)A single image super-resolution reconstruction(SISR)method via introducing multi-denoising autoencoding priors is proposed.In view of the existing SISR methods based on end-to-end mapping often need to train different networks for specific scale factors,we use the denoising autoencoder networks as natural image priors.At the same time,inspired by the average/aggregation idea,we use multi-denoising autoencoding priors to enhance the stability of the method and integrate them into the iterative optimization framework to avoid falling into local solutions.Compared with several other advanced image restoration methods,our proposed image super-resolution method model performs better in terms of PSNRs and visual comparisons.On the whole,the proposed methods make effective use of abundant prior information of images as much as possible,which reflects the importance of priors to image restoration technology.At the same time,the proposed methods also achieve better image restoration results.