RAW Image Super-resolution Method With Deep Learning

High-resolution images not only bring people a more comfortable visual experience in the era of digital media,but also play a role that cannot be ignored in real-world scenes such as security and medical field.However,it is difficult to obtain high-resolution images directly in these scenes due to the limitation of many factors.Therefore,the single image super-resolution(SISR)has been a hotspot in the digital image processing and computer vision.Existing SISR methods are mainly designed for synthetic data,where the manually defined degradation kernel is simpler than that of real data.Thus,the made assumptions limit the performance of these methods on real data,it is also hard to improve generalization by feeding an external degradation kernel.Recently,some works start to explore the SISR problem on real data and have pushed this field forward by several progresses.However,these methods work in the RGB-to-RGB or RAW-to-RGB domain,which does not fully take advantage of the RAW format.Based on deep learning technology,this thesis studies the task of degradation kernel estimation,and futher studies how to better apply the information of degradation kernel to help reconstruct high-resolution images.To tackle the problem that the manually defined degradation kernel is too simple,this thesis proposes a blur kernel estimation method based on minimization of image degradation equation.However,this equation is a highly ill-posed problem since only lowresolution image is known.This thesis makes it possible to minimize by introducing the subspace constraint of the blur kernel.Additionally,the degradation model between HRLR images is closer to ideal distributions in RAW domain,because RAW images are not affected by the image signal processing pipeline,which motivates this thesis to design a blur kernel estimation method based on deep learning and minimization of image degradation equation,and can obtain more accurate results of blur kernel estimation.Based on the advantage of RAW data and the observation that the degradation kernels are closer to ideal distributions in RAW domain,this thesis proposes the first SISR method in the RAW-to-RAW domain for real data.For the highly ill-posed problem if the degradation equation,this thesis extends the subspace constraints to two unkonwns,the blur kernel and the high-resolution image,which enables the minimization of the degradation equation to jointly estimate the blur kernel and restore the high-resolution image.The subspace constraints are learned by two convolutional neural networks,which reformulates an ill-posed SISR problem into a better posed one.In this thesis,the two proposed methods are fully compared with the existing methods on the real datasets.Both the quantitative and the qualitative results indicate the proposed method is superior than existing ones.Ablation studies also verify the effectiveness of each individual component,and the efficiency of the method in this thesis is also verified by the running efficiency experiment.