Research On Efficient Image Restoration Methods Based On Deep Learning

High-quality images show better visual effects since their higher pixel density,clearer image content,and richer texture details.However,the quality of images is usually restricted by various objective conditions such as climatic,acquisition equipment,acquisition methods,image storage,and image transmission.Thence,the technology that can restore low-quality images to high-quality images is essential.It is worth noting that the image degradation process is irreversible.Therefore,as an ill-posed problem,reconstructing high-quality images is still a challenging task.Recently,deep learning has greatly promoted the development of image restoration.However,such methods often generate high-quality images by constructing large and complex networks,which will consume more computing resources,spend more execution time,and greatly limit their application scenarios.To address these issues,we solve image restoration from feature enhancement and regular term constraints.Specifically,we pay attention to the potential features,image priors,and external knowledge of the degraded image,and study how to build efficient image restoration methods by fully using them.Meanwhile,we take the tasks of Single Image Super Resolution(SISR)and Single Image Denoising(SID)as the research objects to verify the effectiveness of the proposed methods.In short,we propose three efficient methods,which involve six models.The main contributions are as follows:? Propose an image restoration method based on multi-scale image feature enhancement.Images will show different features under different scale spaces.Based on this discovery,we propose the first multi-scale image feature enhanced image restoration method.Specifically,we propose two multi-scale feature extraction modules and build two efficient models: Multi-Scale Residual Network(MSRN)and Multi-scale Dense Cross Network(MDCN).These two models realize the extraction,interaction,and fusion of multi-scale image features in a single model.Meanwhile,they achieve excellent performance in SISR and SID tasks,respectively.By making full use of the multi-scale image features,the model can learn the mode of enhancing the image from the geometric features,so that can build more efficient and accurate image restoration models.? Propose an image restoration method based on edge prior guidance.Making full use of image priors is essential for image restoration.Among all image priors,the edge prior is one of the most effective prior.However,the existing edge detectors are sensitive to noise and have poor versatility.Therefore,we propose the first deep edge reconstruction network(Edge-Net)that can directly reconstruct image edges from the degraded images.Meanwhile,we propose two edge prior guided models: Soft-edge Assistance Network(Sea Net)and Multi-Level Edge Feature Guided Network(MLEFGN).These two models use the edge prior provided by the Edge-Net to guide image restoration and achieve excellent performance in SISR and SID tasks,respectively.By making full use of image edges,we can further characterize the edge of the reconstructed image and constrain the solution space,so that can build more efficient and accurate image restoration models.? Propose an image restoration method based on knowledge distillation.Due to the limitations of the degraded image and the model structure itself,reconstructing high-quality images with lightweight model is still a challenge.Therefore,we introduce the knowledge distillation mechanism to realize automatic learning of features and priors by using the deep latent knowledge of the model itself or the external knowledge provided by the heterogeneous models,thus further improving the model performance.To achieve this,we propose two efficient models: Progressive Self-Distillation Network(PSDNet)and Heterogeneous Knowledge Distillation Network(HKDNet).The performance of these two models can be further improved by using the introduced external knowledge and they achieve excellent performance in SISR and SID tasks,respectively.By introducing the knowledge distillation strategy,the model can automatically learn rich information and further constrain the solution space,so that can build more efficient and lightweight image restoration models.