Research On Lightweight Image Super-Resolution Algorithm Based On Frequency Division

As all walks of life have fully entered the information age,digital images are one of the main carriers that carry people’s life perception,transmission,processing,and analysis.And the image quality directly affects people’s visual experience.Therefore,the quality of the image is particularly important.However,due to the complex conditions in real life,the images we collected and saved are usually of low resolution,which is difficult to meet people’s demand for increasing resolution.Image super-resolution,as a very classic and popular underlying task in the computer vision field,can not only significantly improve the resolution and perceptual quality of images,but also can be used as preprocessing for other computer vision tasks to significantly improve accuracy.Therefore,the super-resolution task has important practical application value and academic value.Deep learning-based super-resolution algorithms are currently the mainstream and optimal method,but they generally face the problems of high algorithm complexity and huge parameters,which seriously affects the deployment of super-resolution algorithms on low-power devices.Therefore,from the starting point of reducing feature redundancy and improving feature learning efficiency,this thesis designs efficient and lightweight superresolution algorithms,and proposes the following two works:1、A super-resolution algorithm of the extractive-distillation residual network is constructed.The algorithm explores the relationship between the intermediate features,divides the intermediate features into two parts,the features to be recovered and the recovered features.We propose the extraction residual which only retains the recovered features.At the same time,in order to alleviate the problem of feature redundancy calculation,we introduce channel distillation among modules.By compressing the feature quantity,the network can focus on the feature to be recovered without consuming computing resources on the recovered feature.In addition,based on the idea of divide and conquer,we proposed the extraction-distillation residual by combining the extraction residual and channel distillation.Experiments show that our method can effectively reduce redundant features,network parameters,and computational complexity,and its performance is better than the best existing lightweight network.2、A single-image super-resolution algorithm of frequency division residual network is constructed.The algorithm is based on the extraction-distillation residual network and explores a more reasonable scheme to retain the intermediate features.By introducing shared features,frequency division residual models a more reasonable intermediate feature relationship explicitly.In addition,the difficulty of network restoration is different for different frequency features,and most existing networks deal with all frequency features indiscriminately,resulting in the problem of low efficiency of feature learning.We propose a feature frequency division learning strategy.By dividing the intermediate features into multiple frequency band features,we adaptively allocate the network depth required by different frequencies,which effectively improves the efficiency of network feature learning.Compared with the extractive distillation residual network,the model complexity of the fractional residual network is reduced by about 9%,the performance is further improved,and a better tradeoff between performance and complexity is achieved.