Research On Image Super-resolution Algorithms Based On Deep Learning

In recent years,with the widespread application of the concept of globalization,computer technology has flourished,and image information has played a pivotal role in many information transmission processes.Due to the different factors(such as cost)in different scenes,images acquired in daily life and special environments such as video surveillance,medical imaging,and remote sensing satellites is generally low resolution.In the low-resolution image,due to the lack of high-frequency information,the detailed information of the image can not be clearly distinguished.Therefore,the superresolution reconstruction technique,converting the low-resolution image into a highresolution image using signal processing and image processing methods,has been paid more attention and widely used,this technique has become one of the important topics in computer vision.The main task of the image super-resolution algorithm is to reconstruct the highfrequency information in the high-resolution image,which can clearly depict the image details,and provide a clear image source for many computer vision tasks such as target recognition,target tracking,etc.There are important applications in the fields of medical imaging and video surveillance.In recent years,the learning-based super-resolution reconstruction method has received extensive attention.This method uses a large number of sample training to find mapping relationships in low-resolution and high-resolution images to reconstruct high-resolution images.Deep learning as one of the learning algorithms,because of the high quality of reconstruction and the no need to manually extract features,has become a research hotspot.The main research point of this paper is the super-resolution reconstruction algorithm based on deep learning.(1)For the characteristics of image super-resolution reconstruction,constructing the dense residual network structure by using the advantages of convolutional neural network in image processing,altering the residual network structure and combining with dense network structure.The complement 0 operation is performed around the image to keep the feature images size.(2)Based on the dense residual network,building the recurrent redisual dense network by combining with the recurrent network.More residual layers are used in the network structure,and more dense shortcut connections are used to make full use of the inter-layer feature maps.The combination of the recurrent network in the network further deepens the number of network layers and reduces the complexity of the parameters.In this paper,a large number of public datasets are used to train and verify.We use the PSNR and SSIM to evaluate the reconstructed images,and the results verify the effectiveness of the now networks in image super-resolution reconstruction.In the terms of vesion,the reconstructed images also have a good visual effect.