Application And Research Of Deep Convolutional Neural Network In Image Restoration

Image super-resolution is an important research area,and it is also seen as a key technology in the field of image restoration.In most of electronic imaging applications,we need to collect and analyze high-resolution images.High-resolution images mean higher pixel density in the image,so these high-resolution images can provide more critical details in various applications.However,due to the limitation of hardware equipment,harsh environment and many other factors,the quality of the images collected is bad,which greatly limits the quality and safety of the work.Therefore,we need to put forward reliable ways to repair and improve the quality of the collected images.The research of single image super-resolution restoration is of great significance.At present,the research of image super-resolution has made good progress.Interpolation-based and reconstruction-based image super-resolution methods have low complexity,but the performances are limited.These methods are difficult to repair the high-frequency details of the image.Due to the vigorous development of artificial intelligence,deep learning technology has became more and more popular in the field of computer vision,and has made remarkable achievements.Therefore,image superresolution based on neural network has became a research hotspot.In addition,the image super-resolution model based on neural network can also be used in other areas of image restoration,such as image denoising,image restoration and so on.In this paper,the image super-resolution based on neural network methods are studied as the research masterstroke.In order to further improve the quality of the restored image and solve image super-resolution high-frequency details loss,edge blurring and so on,this thesis proposes some new methods.The main researches are as follows:(1)A dense many-to-many connections deep convolutional neural network model is proposed to overcome the limitation of learning ability of conventional learningbased image super-resolution restoration methods,which leads to poor restoration and reconstruction results.In the feature extraction stage,the model uses many-to-many connections for different layers.These connections can effectively solve the gradient disappearance and increase the transmission of characteristic information.In addition,some parallel convolutional kernels which size are 1?1 is used in the image reconstruction stage,which enhances the nonlinear representation of the whole model and improves the expression ability of the whole network.(2)Aiming at the problem that the effect of edge detail recovery is limited by existing models,a multi-viewfield parallel convolutional neural networks are proposed.The model uses convolutional kernels of different sizes(1?1,1?1,3?3),which can ensure that images with different sparseness have reliable perception areas.In order to reduce the computational complexity of large convolutional kernels,we put convolutional kernels of different sizes in a parallel position,allowing them to operate in parallel,thus ensuring a faster speed.