Research On Image Super Resolution Technology Based On Deep Learning

Research on image super-resolution based on deep learning mostly focuses on the design of super-resolution network models.This type of model assumes that low-resolution(LR)images are down sampled version of high-resolution(HR)images obtained by using bicubic interpolation.Although the peak signal-to-noise ratio is continuously improved on the validation set that is artificially generated in the above way,since the degradation methods of real LR images are various and do not meet the assumptions of the model,the effectiveness of the model will be greatly reduced.SRMD(Super-Resolution network for Multiple Degradations)proposed a super-resolution model that solves multiple degradation problems by modeling kernels and noise.The disadvantage is that when the image blur kernel is unknown,the network has to use grid search strategy to find a suitable Gaussian blur kernel to process the image;hence the method is difficult to be applied to the real scenario.It can be seen that an efficient kernel estimation method can help the super-resolution model to be better applied in real degraded image scenes.To this end,this paper proposes a kernel estimation model based on the spectrum.In this method a convolutional neural network is used to learn its own kernel information from the spectrum of the degraded image.That is,by modeling the key factor of image degradation-blur kernel,a network model that can directly predict the blur kernel from the degraded image itself is constructed.The kernel predicted by this model can be directly used in the existing super-resolution model to make it better applied in the real scenarios.On the other hand,since the corresponding HR image of the real degraded LR image is not available,the real degraded LR image cannot be directly used in training.Therefore,the proposed kernel estimation model can be used to extract the kernel information of different degradation methods in the LR dataset.The extracted kernel is then used to perform blur operations on the HR dataset to generate a paired data set that conforms to the real-world degradation method.The generated paired data are thus used to train the super resolution network.The image-specific GAN(Generative Adversarial Nets)network can be used in the kernel estimation model,which is based on the unsupervised method kernel GAN(Kernel Estimation using an Internal-GAN),to predict the blur kernel of LR images.However,there is a problem that the network cannot converge stably.Enlightened by the ideas of High-to-Low GAN and DSGAN(Down Sample GAN),this research proposes an image-specific GAN network model based on Kernel GAN to simulate the degradation of specific LR images,by making use of the information from HR image.In this method,a frequency separation mechanism is introduced.By employing a high-pass filter to the discriminator,the discriminator can more easily focus on high-frequency image features,effectively speeding up the network training process.In this research,the output of various kernel estimation methods is compared on the DIV2KRK(DIVerse 2K Resolution Image Dataset with Random Kernel)testset.Subjectively,the accuracy of the kernel estimation model we proposed is significantly higher than other existing methods.To objectively verify the effectiveness of blur kernel estimation,we applied blur kernel estimation to the existing super-resolution model,and compared the results of different superresolution methods on synthetic datasets and real degraded images.Our model has achieved competitive results on the objective evaluation indicators PSNR(Peak Signal-to-Noise Ratio)and SSIM(Structural Similarity Index).