Degradation Modeling And Algorithm Design For Real-world Image Restoration

Image restoration is an important research topic in low-level computer vision and has wide applications in daily applications,surveillance tasks,and astronomical re-searches.Recently,with the development of deep learning,algorithm for image de-noising has seen great progress.Generally,a pre-defined degradation model is required to generate training samples in pair.In this training paradigm,several kinds of charac-teristics and mechanisms are adopted to boost the performance of image restoration.In this thesis,we study image restoration from three perspectives:algorithms,data-based,and task-based real-world degradation modeling.Details are listed below:Deep boosting for image denoising and its applications in real-world scenarios.We propose a deep boosting algorithm for real-world image denoising.It replaces conven-tional boosting units by elaborated dense dilated fusion network,which addresses the vanishing of gradients during training due to the cascading of networks while promot-ing the efficiency of limited parameters.To facilitate leaning-based methods including ours,we build a new real-world image denoising dataset.Based on the dataset,we conduct comprehensive analyses on the domain shift issue for real-world denoising and propose an effective one-shot domain transfer scheme to address this issue.Antiforensics-driven degradation modeling and denoising method.Camera trace is a unique noise produced in digital imaging process.Most existing forensic methods analyze camera trace to identify image origins.In this thesis,we address a new low-level vision problem,camera trace erasing,to achieve the adversarial attack for trace-based forensic methods.To effectively erase camera trace while avoiding the destruction of content signal,we propose to model this unique noise in a forensics-driven manner.Based on it,we propose Siamese trace erasing,in which a novel hybrid loss is designed on the basis of Siamese architecture for network training.Compared with existing anti-forensic methods,our proposed method has a clear advantage for camera trace erasing,which is demonstrated in classification,clustering,and verification tasks.Degradation modeling for real-world image super resolution and a light-weight method.We investigate SR from the perspective of camera lenses,named as CameraSR,which aims to alleviate the intrinsic tradeoff between resolution(R)and field-of-view(V)in realistic imaging systems.Specifically,we view the R-V degradation as a latent model in the SR process and learn to reverse it with captured low-and high-resolution image pairs.To obtain the paired images,we propose two novel data acquisition strate-gies for two representative imaging systems(i.e.,DSLR and smartphone cameras),re-spectively.Based on the obtained dataset,we quantitatively analyze the performance of commonly-used synthetic degradation models,and demonstrate the superiority of Cam-eraSR as a practical solution to boost the performance of existing SR methods.In addi-tion,we propose a light-weight SR method,which combines the transposed convolution and spatial aggregation.It achieves real-time SR without performance degradation.In summary,we focus on the degradation modeling and algorithm design for real-world image restoration.Compared with previous works,we demonstrate the superior performance of proposed algorithms on public or newly constructed datasets.