Image Restoration Based On Deep Learning

Image restoration is one of the core issues in the field of computer vision,aiming to reconstruct high quality image from low-resolution input.High quality images are playing an increasingly important role in modern life.However,due to the limitation of image acquisition equipment and loss during transmission process,the resolution of most images can not satisfy the requirements in real applications.However,the existing image restoration methods have poor generalization ability in real scenes.In order to achieve this goal,this work firstly studies about image restoration under the restricted conditions with known degradation types.Then we extends the image restoration problem to real scenes under unrestricted conditions with unknown degradation types.This study mainly achieves the following research results:1.In order to improve the quality of super-resolution image with known degradation,this thesis proposes Enhanced SRGAN for enhancing image details and face super-resolution(USRN-SFT)algorithm for improving the accuracy of face images respectively.This work studies the network architecture,adversarial loss and perceptual loss,so that the proposed ESRGAN not only improves the model capacity and reduces the taining difficulty,but also improves the visual quality of reconstructed images.In this work,after a fair comparison and analysis of existing face algorithms,a more effective face prior utilization and network structure is proposed.USRN-SFT can reconstruct more accurate results compared with existing face algorithms.2.In order to improve the quality of denoised image with known degradation,this thesis proposes RDS-Denoiser for improving quality of reconstucted image.This work firstly synthesizes a large amount of training data to approximate the real data.Then,this study decomposes the denosing process into two stages.The Sage-I Denosing is to predict the noisy map of noisy image.The Stage-II Denosing to further improve the visual quality.Experiments show that RDS-Denoiser achieves competitive performance comparing to state-of-the-art denoising methods.Since the effectiveness of the model depends heavily on the matching degree between the synthetic data distribution and the test data,the next step is explored.3.In order to improve the quality of restored image with unknown degradation in real scenes,this thesis proposes data constraint strategies to improve the current Cycle GAN based framework in three aspects,including loss function,training data and post-processing.We term the new method as Enhanced Cycle GAN(ECycle GAN),which is dedicated to image restoration task.First,we introduce a new image constraint loss function to compensate the absence of pixel-level supervision in unsupervised learning.Second,we constrain data content for discriminator,which encourages the discriminator to suppress high-frequency textures or artifacts.Third,to further address the model selection issue and preserve data content after training,we introduce a model average strategy for post-processing.Benefiting these improvements,the proposed method ECyle GAN can not only stabilize the training process but also improve the performance of image restoration.4.In order to improve the quality of single restored image with unknown degradation in real scenes,we conduct experiments on the input-self.Inspired by the recent success of single image generation based method Sin GAN,we tackle this challenging problem with a refined model SR-Sin GAN.First,we empirically find that the introduced input prior can improve the robustness of the generation model.Second,we introduce Global Contextual Prior to provide semantic information.Finally,we design an image gradient based local contextual prior to guide detail generation.Experimental results show that the introduced contextual priors can stabilize the training process and preserve the fidelity of outputs to improve the generated image quality.