The Research On Image Super-resolution Reconstruction Algorithm Based On Generative Adversarial Network

With the rapid development of the information age,people's demand for image quality is increasing.Image Super-Resolution(SR)reconstruction aims to recover natural and clear textures from low-resolution images through computer software,and reconstruct high-resolution images with good quality.In recent years,thanks to the application of deep learning technology,image SR reconstruction has made significant progress,and it has good applications in consumer and medical,public security,military and other related professional fields.Aiming at the problems of slow training speed of image SR reconstruction algorithm,unstable network model and fuzzy high-frequency information processing,this paper has carried out a lot of research work on single image SR reconstruction task,and proposed a novel image SR reconstruction method.The main contents are as follows:(1)Through in-depth exploration of the relevant research of the existing image SR technology,and analysis of the many advantages and disadvantages of the previous research results,this article systematically explains the relevant theoretical basis of image SR reconstruction.In order to solve these problems in the existing algorithms,We proposed a Novel image SR reconstruction method.(2)Aiming at the defects of traditional SR reconstruction algorithm such as unstable model training,large parameter amount,and complicated calculation,this paper proposes an image SR reconstruction algorithm M?SRGAN based on multi-scale GAN.Nowadays,Lots of existing algorithms generated artifacts in images,We propose a multi-scale residual dense connection block as the basic building block of M?SRGAN to improve the quality of the reconstructed image;the traditional SR reconstruction algorithm is susceptible to noise and causes the problem of ringing effect in the reconstruction result.In M?SRGAN,a new long-short jump connection residual learning method is introduced to improve training effects and maximize information exchange;The generation network removes the BN layer,retains the objective contrast information of the image,improves the visual effect of the image and saves computing resources.(3)The loss function of the traditional algorithm blurs the details of the image during the training process.By reference SRGAN,this paper constructs a loss function with texture difference constraints to guide the generator to focus on the image texture details.Rebuild images with realistic texture details.(4)In addition,in view of the existing deep learning-based methods that explore the global features of the image on a global scale,they ignore the correlation between the image features and the high-frequency feature information of the image,which leads to problems such as the smoothness of the reconstructed image.An improved GAN image SR reconstruction algorithm based on the attention mechanism,which realizes the weighted enhancement of features,adaptively recalibrates the feature response in the channel,screens out features that are more conducive to reconstruction,and improves the model Training rate.Through comparative analysis of test experiments on several classic test data sets,the results prove that the method proposed in this paper improves the stability,timeliness and accuracy of the image SR reconstruction algorithm,and its comprehensive performance is relatively prominent,which has good research significance and practicality value.