Design And Implementation Of Super-resolution Based On Deep Learning

Super-resolution can enhance digital image quality at software levels,thus breaking the limitation of technical level of image-forming system.It has extensive applications in medical imaging,satellite imaging,surveillance and other fields.Therefore,it is a hot issue in field of computer vision.In recent years,the deep learning and generative adversarial networks has been used in super-resolution,which made a great breakthrough in super-resolution problem.Then many super-resolution algorithms which perform well beyond traditional algorithm have been proposed.Firstly,this paper reviews the development of super-resolution and described the research situations of super-resolution.Then we analysis the theoretical model of superresolution and introduce the classical super-resolution algorithms.Finally,based on the one-to-one correspondence between reconstructed image and high-resolution image in super-resolution problem,we propose an improved super-resolution generative adversarial networks based on siamese discriminator.The main improvement of this paper are as follows:This paper proposed the new structure of siamese discriminator,which change the limitations that the discriminator of native generative adversarial networks can only inputs a single sample.The siamese discriminator can receive image pairs consisting of one-to-one reconstructed image and high-resolution image.It can take advantage of the additional information without increasing the scale of the network.This paper improved the perceptual loss based on pre-trained VGG network and proposes the dynamic perceptual loss.Experiments show that the improved dynamic perceptual loss can compare the difference between reconstructed image and highresolution image in high-level features.The dynamic perceptual loss will be optimized in the whole adversarial training.And the dynamic perceptual feature will gradually close to the feature related to the image clarity.Using the dynamic perceptual loss,the generator can focus on the features highly related clarity.Which means the generator can reconstruct the image which is closer to the high-resolution image.In order to avoid the defect of mean opinion score,this paper proposes an improved image subjective quality evaluation method based on relative opinion,which is named as mean relative opinion score.The MROS reduce the rank of score and set the baseline of image quality,which can reduce the instability of scoring process.This paper compared the performance of SRCNN,FSRCN,ESPCN,SRGAN and SDSRGAN in different super-resolution dataset.In the experiments,this paper uses the PNSR,SSIIM and the subjective evaluation method which proposed in this paper to compare the quality of reconstructed image which generated by different super-resolution algorithms.