Research On Image Super-resolution Algorithm Based On Deep Learning

Super-resolution reconstruction technology restores the missing high-frequency information in low-resolution images through algorithms to reconstruct high-resolution images.The super-resolution reconstruction method based on deep learning can automatically extract the features of the image and avoid the problem of missing information due to manual selection of features.However,new problems such as the disappearance of the gradient and the unstable model training have appeared,and the problem of distortion will occur in the reconstruction process of the image with colorful colors.In order to solve the above problems and improve the quality of the reconstructed image,this paper studies the optimization method of attention mechanism and generative adversarial network in the super-resolution field,and proposes two new super-resolution reconstruction algorithms.The main work of this paper is as follows:(1)A super-resolution reconstruction algorithm(RAMSR)based on residual attention mechanism is proposed.First,a cascaded residual network of global residuals and local residuals is constructed,which can not only fully mine the internal features of the image,but also alleviate the problem of network degradation;Then,the improved channel attention mechanism network is added to the residual network to enhance the ability of the neural network to extract high-frequency features;Finally,a deep separable convolutional network is introduced before the channel attention mechanism unit to reduce model parameters and improve parameter utilization.Experimental results show that compared with methods such as SRCNN,VDSR,EDSR,and RCAN,RAMSR algorithm improves the quality of reconstructed images,reduces model parameters,and reduces training time under the same conditions.(2)A generative adversarial network algorithm(RHAMGAN)based on residual mixed attention mechanism is proposed.On the basis of SRGAN,the generator network and the discriminator network are improved.The first is to design a generator network based on the residual dense network in the generator network,and add an improved hybrid attention mechanism to it,so that the neural network pays more attention to the high-frequency information of the image,And makes the image generated by the generator network more similar to the original image;Second,in the discriminator network,the relative discriminator network is used to replace the standard discriminator network,so that the neural network outputs the probability that the real image is more real than the fake image,making the learned texture more detailed.The experimental results show that,compared with the SRGAN method,the RHAMGAN algorithm can improve the PSNR and SSIM averages,reduce artifacts,and have a good image reconstruction effect and a stronger sense of reality.