Research On Single Image Super-Resolution Reconstruction Algorithm

At present,the hardware equipment and transmission bandwidth are limited,and the image resolution quality collected by the terminal is difficult to reach the required level,so it can not meet the actual application requirements.How to use low-resolution image to reconstruct its corresponding high-resolution image and restore the high-resolution image with richer details and clearer texture has become an urgent need in many fields such as video surveillance,satellite remote sensing,medical image and so on.It is more difficult to study single image super-resolution reconstruction than multiframe image,especially high-multiple reconstruction.With the extensive research of artificial intelligence and in-depth learning,image super-resolution technology based on learning has become the current research hotspot.Existing methods based on in-depth learning use specific network structure.When the number of network layers is deep,it is difficult to converge when training the network,and the effect of superresolution is affected by the performance of the network itself,thus affecting the image quality and effect of network reconstruction.In this paper,an improved residual network reconstruction model based on channel attention is proposed,which enables the network to learn the important features of the channel as much as possible during training.At the same time,a multi-model fusion reconstruction framework based on texture feature classification is proposed.The network with different texture features is trained separately,so that it can focus on the types of image texture that it is good at,thus improving the network.Index and effect of super-resolution reconstruction.Firstly,this paper introduces the research status and challenges of single image super-resolution reconstruction.Secondly,it introduces the current methods and classification of image super-resolution reconstruction,and the theoretical basis of image super-resolution reconstruction based on residual network.Secondly,it introduces in detail some improved methods based on depth residual network proposed in this paper,and gives the improved network structure model and related theoretical derivation.Then,the methods of texture feature extraction and K-means clustering multiple selection network are introduced in detail The proposed multi-model fusion reconstruction network based on texture feature classification,compared with the current advanced methods in common data sets,has significantly improved both in subjective effect and objective evaluation index,thus verifying the effectiveness of this method.