Research On Super-resolution Reconstruction Method Based On Deep Learning

In the era of rapid development of information alternations,images as the main carrier of communication information,are applied to various scenes that are familiar to the public and necessary,such as medical and military fields,satellite monitoring fields,digital media and many other fields,which have higher requirements on image quality.General low resolution image is difficult to be accepted by the public,people need more colorful images to meet the visual enjoyment.Therefore,image processing and computer vision have become important applications of super-resolution(SR)reconstruction technology.In recent years,deep learning technology has made breakthroughs in many fields,and the effect of convolutional neural network and generative antagonistic network in image reconstruction is very significant.Therefore,combining with the theory of deep learning,this paper proposes an image super resolution algorithm based on the convolution neural network and an image super resolution algorithm based on the condition of generating antagonistic network.The main work content is as follows:1.An image super resolution algorithm based on convolutional neural network is proposed.Combining the characteristics of the convolutional neural network model and residual network structure,this paper proposes a new deep network structure for image super-resolution reconstruction.The basic idea is to use four SRCNNS as the basic structure to build two continuous residual learning networks.And the method use the residual network to learn the residual components of high frequency that can not be recovered by conventional super resolution methods.Compared to the reconstruction of the traditional method,the method solve the convolution image super-resolution method of the neural network to a certain extent.For example,poor robust performance,complex network parameters,reconstruction after the detail texture is not clear and other questiones.The SSIM and PSNR has a significant improvement.It is proved that the method can recover more image detail information.2.An image super resolution algorithm based on conditional generative antagonisticnetwork is proposed.The basic idea is to make use of the structural characteristics of the generated antagonistic network,the generator first generates the "forged" data sample,and then the discriminator determines the authenticity of the generated sample.When the generated sample is basically consistent with the real sample,the discriminator will believe it to be true,and finally complete the training of the generated antagonistic network.The algorithm is compared with the traditional image super-resolution algorithm,experimental results show that the algorithm on the peak signal-to-noise ratio and structural similarity significantly increased,but the discriminant essence as the discriminator,about the authenticity of the input samples,so in terms of subjective visual,generator "fake" data sample,makes the algorithm reconstruction effect more realistic.