Research On Image Restoration Algorithm Based On Deep Convolutional Network

With the popularization and development of imaging technology,digital images have gradually become an important information transmission medium.However,in actual application scenarios,due to imaging conditions and external interference and other factors,there are quality degradation problems such as low resolution and missing data.Image restoration aims to study how to recover ideal images from degraded images to achieve improvement.The purpose of image quality.The deep convolutional neural network can effectively obtain the hierarchical feature representation of the image through the hierarchical processing of the convolution operation,which brings a new modeling method to the image restoration,and brings the effective improvement of the performance of the image restoration algorithm,which makes the research progress A substantial breakthrough.For this reason,this thesis mainly studies the single image restoration algorithm based on deep convolutional network,and proposes corresponding improved images for the two situations of low image resolution and missing images obtained in daily applications.Quality methods,specific results include:(1)An image super-resolution attention network driven by perceptual loss is proposed.Unlike common supervised learning,which requires paired data set support and network pre-training,this method only needs low-resolution images that the task depends on to complete the task,which solves the waste of time and space resources.On this basis,in order to further strengthen the network's ability to recover images with excellent quality,the spatial residual attention module and the use of perceptual loss to guide network training are added to the network,which also makes the images recovered by this method have excellent images Visual effects.(2)A two-channel feature equalization image patching confrontation network is proposed.In this thesis,we use the adversarial learning method to achieve image deletion modification,and design a generation network with a codec structure.While ensuring that it reshapes the detail texture part,it uses dual-channel feature equalization to strengthen the network's reconstruction of the image structure,and through the relativistic global and local discriminator further restricts the consistency of the restored image.Experimental results show that this method can recover images with reasonable structure and rich details.