Image Super-resolution Reconstruction Based On Deep Learning

With the development of science and technology,higher requirements on the image resolution have been put forward to image analysis,understanding and recognition.But there are still some problems to capture an image at a desired high-resolution level.Therefore,how to obtain high resolution images and videos has become one of hot research problems in the field of image processing.Super-resolution technology has been widely used in the fields of object detection,image coding,satellite remote sensing,video monitoring and medical diagnosis and so on.Algorithms of image resolution includes interpolation based methods,reconstruction methods,learning based methods,deep learning based methods.Traditional methods need manually extract features.There is a limitation.Deep learning based methods could automatically extract features,getting more and more attention.But there are still some insufficient,such as need enlarge preprocessing,single size of kernel size,single channel,can't meet the demand of the human eye to watch,etc.Therefore,in order to solve the above problems,we propose two kinds of structures of deep learning network for single image super-resolution reconstruction.First,we propose a shallow and deep convolutional neural network of two channels for the single image super-resolution.Firstly,the shallow channel mainly restores the general outline of the image.On the contrast,the deep channel extracts the detailed texture information.The two complement each other very well.Secondly,the proposed method directly learns an end-to-end mapping between low-resolution and highresolution images.The upsampling of the network by deconvolution is embedded in the two channels,so it doesn't need hand-designed preprocessing.Finally,during the last period of reconstruction,the deep channel adopts multi-scale manner,which can extract both the short-and long-scale texture information simultaneously.Our model is evaluated on three different public datasets including images and videos.Average values of PSNR on three datasets higher than the existing algorithms.The value of PSNR for the reconstructed lenna image expanded 3 times is up to 39.97 dB.And the details are rich,the texture is clear.A size of 280 * 280 image reconstruction requires only about 0.17 s.Experimental results demonstrate that the proposed method outperforms the existing methods in accuracy and visual impression.Second,Algorithms adopt mean square error as loss function,which is easily to make the super-resolution image too smooth.Generative adversarial network for image-resolution is to solve the above problem,it fuses mean square error and human eye perception to the loss function.In this paper,we put forward the modified generative adversarial network.Sub-pixel convolution is adopted to upsampling,which is more likely to find the connection between the adjacent pixels.Because this step does not have a convolution operation,the network computes faster.The super resolution image can meet the demand of the human eye to watch.Experiments use lenna image test,the expanded twice after rebuilding image PSNR is only 29.20 dB.But the detail of reconstruction image is rich,meeting the human visual perception.