Research On Super Resolution Image Based On Deep Learning

With the development of multimedia technology,people will receive a lot of image information every day.The higher the resolution of the image,the stronger the expression ability of detail features,the clearer the image.In practical application,due to the limitation of objective factors,the image resolution may be relatively low,so it is necessary to improve the image resolution.Image super resolution is the process of reconstructing a high-resolution image from a given low resolution image.At present,deep learning is widely used in image super-resolution reconstruction.It can automatically extract the relevant details,so as to establish the relationship between low resolution and high resolution images.Compared with the traditional algorithm,the reconstruction effect is better,but this kind of algorithm still has some defects,and there is still much room to improve the image quality and speed.Therefore,in this paper,based on deep learning,the image super-resolution algorithm is deeply studied.The main purpose is to add the attention mechanism to the existing classical deep learning model,and propose an improved optimization model.The work done in this paper is as follows:(1)The current mainstream super-resolution algorithm based on deep learning is studied.After comparative analysis,it is decided to optimize the algorithm by referring to the classic super-resolution algorithm VDSR and srgan model.(2)Based on the VDSR network model,an optimization model is proposed.Aiming at the problem that the image details are not rich enough after VDSR reconstruction,ECA channel attention module is introduced to acquire more image feature information by learning the correlation between the channels,which increases more image details,and thus achieves better reconstruction effect;The residual learning module of VDSR is optimized,which reduces the total parameters to a certain extent and improves the network running speed.(3)Based on srgan network model,an optimization model is proposed.Because batch normalization layer often ignores some image details in super-resolution image reconstruction and increases network complexity,this paper removes BN layer in srgan generator,and introduces ECA channel attention to obtain corresponding weight for each feature map generated by residual block,and can process more image details,This can not only improve the speed of network model,but also reconstruct the image with better visual effect.In order to evaluate the effect of super-resolution reconstruction,PSNR and SSIM are used as objective evaluation indexes,and subjective evaluation is combined.After open data set training and many comparative experiments,the results show that the two improved models proposed in this paper are more abundant in detail recovery and faster in model processing than the contrast model,The evaluation index is better.