Deep Learning-based Methods For Lightweight Image Super-resolution

With the rapid popularity of Tik Tok-style applications(apps)and the rise of streaming media platforms,video or image has become an important medium for people's daily communication,and users have increasingly higher requirements for image quality.To solve the degradation of image quality caused by hardware devices or network bandwidth and improve the image quality of mobile devices or web apps,Image Super-Resolution(SR)algorithms are commonly used.The goal of SR is to estimate the high-resolution image from the given low-resolution counterpart.Thus,SR technology is widely used to increase the spatial resolution of the image and enhance image quality.Furthermore,it is still a challenging task due to the ill-posed nature caused by the degradation process.Thus,Image SR is a classical and fundamental problem in low-level vision.Recently,most existing deep learning-based single image super-resolution(SISR)algorithms rely on increasing the complexity of the model to improve the performance of image super-resolution reconstruction.However,due to the high complexity of these models and the rising cost of computing resources,many advanced algorithms are difficult to deploy in actual application scenarios.In order to strive towards the end goal of deploying SISR models for real-world apps,this work presents lightweight image super-resolution solutions to solve this problem,in which SR algorithms work well under resource-constrained conditions.In summary,the main research contents are as follows:(1)We propose a lightweight SISR model based on attention mechanism and multiscale residual module.First,the dense connection and channel attention mechanism are used to enhance the multi-scale representation ability of the proposed model,thereby improving the super-resolution reconstruction performance.Second,in the training phase,the mean square error loss and the total variation norm are combined as the loss function to remove the noise in the final SR result.Finally,through extensive experimental analyses and comparisons of results,it is concluded that the proposed SISR method has better reconstruction performance and faster processing speed.(2)We introduce a model based on global residual learning,feature weighted fusion,and multi-scale mechanism to achieve efficient SR.In this work,we first analyze the limitations of the existing lightweight image super-resolution methods.Then,we propose a depthwise separable convolution-based multi-scale residual model for efficient SR and prove the effectiveness of the method through a series of ablation experiments.Lastly,quantitative and qualitative comparisons with several state-of-the-art approaches are conducted on five benchmarks and real low-resolution images.Extensive experiments show that the given model utilizes only a modest number of parameters and operations to achieve competitive SR performance on different benchmarks with different upscaling factors.