Image Super-resolution Using Light-weight Deep Learning Methods

With the development of hardware equipment and multimedia technology,mobile phones,computers and large-screen TVs have gradually formed diversified home display devices.Considering the different computing capabilities and image resolution requirements of different devices,the main content of this paper is to design corresponding lightweight super-resolution network models for different home display devices,which are applied to complete the offline super-resolution of lowresolution images and videos.Different home display devices have different computing capabilities and requirements of image resolution.Considering the limited computing resources of home display devices,the main research goal of this paper is to design lightweight super-resolution networks according to different home display devices.The main work of this paper can be divided into two parts:1.For large-scale super-resolution tasks in desktop device,this paper proposes a recursive multi-stage upscaling network(RMUN)based on the Laplacian pyramid.The RMUN has three innovations:(1)A lightweight feature extraction module(FEM)is proposed,which can fully integrate local deep and shallow features,strengthen the fluidity of features in the network.The FEM is simple in design and has fewer parameters,which reduces parameters and calculations of the model.(2)Multiple subupscaling modules(SUMs)are proposed,which means that the output features of each feature extraction module are up-scaled,increasing the expressive power of shallow feature information in the network.Besides,multiple up-scaling branches also provide multiple error feedback paths,which reduces the difficulty of model training.(3)A discriminative selfensemble module(SEM)is proposed,which discriminately fuses the highresolution features from multiple up-scaling branches,and improves the performance of the model.Experimental results show that the recursive multi-stage upscaling network proposed in this paper has achieved good results in both quantitative and qualitative analysis.On the public test data set BSD100,the PSNR of × 4 super-resolution reaches 24.66dB,which is better than other mainstream algorithms.In addition,it has achieved a good balance between model parameters and model performance.2.For mobile devices with limited computing resources,this paper proposes a multi-level supervised compact network(MSCN)based on the knowledge distillation strategy.The main innovations of the network are:(1)The depth-wise separable convolution is applied to redesign a lightweight super-resolution network structure;(2)The knowledge distillation strategy is utilized to train the light weight network.MSCN employs a compact model(student model)supervised by the HR label,but also supervised by the intermediate representation of a cumbersome model(teacher model).As the experiments show,the parameter of the MSCN is only 0.20M,which is easy to be deployed on home mobile devices.Furthermore,the performance of the student model after knowledge distillation training is better than that of the student model without knowledge distillation training,which verifies the effectiveness of the multi-level supervised compact network.The knowledge distillation strategy can improve the performance of light weight models.