Research On Lightweight Super-resolution Algorithm Based On Non-local Information Fusion

In order to obtain high-resolution images with better subjective effect under limited conditions,super-resolution algorithms are often used to process images to restore image details and improve image quality.With the wide application and development of deep learning methods,convolution neural network is especially suitable for image tasks,and there are many related studies in super-resolution algorithms.However,some super-resolution algorithms with excellent performance based on deep learning have a large number of parameters and a large amount of computation,so it is difficult to apply in the actual scene,so some lightweight models are proposed to solve this problem.This paper studies some problems existing in the current super-resolution algorithms.In order to reduce the number of parameters and the amount of computation,many lightweight networks will reduce the size of the main channel of the model,compress the number of channels in the network,reduce the feature dimension,worsen the nonlinear mapping ability,and affect the performance of the model.In this paper,a lightweight convolution layer based on 3D convolution is designed,which ensures a large feature dimension when the number of parameters is small.According to the proposed convolution layer characteristics,the network feature extraction,feature mapping and image reconstruction are improved,and a better performance is obtained in the case of the number of small parameters.Finally,the comparative experiment shows that the performance of the proposed model is the best in the model with similar number of parameters,and the number of parameters and the amount of calculation are very balanced.In the supplementary ablation experiment,the influence of different network hyper-parameters on the performance of the model is discussed.There are many cross-scale similar image blocks in the natural image,which can be used to supplement the lost information of down-sampling;at the same time,due to the limitation of the calculation method of the convolution layer itself,it is often unable to obtain enough receptive field and make use of the information of the distant position.In order to make full use of the distant and cross-scale similar information in the image,a cross-scale non-local information fusion block is designed in this paper.Because non-local algorithms often have a large amount of computation and are not suitable for lightweight networks directly,this paper uses local sensitive hashing to reduce the amount of computation in cross-scale similar feature search.at the same time,crossscale similar feature information is used in the module,which can more effectively help the network to carry out image reconstruction,better supplement the information lost in the process of image degradation,and obtain better performance.The final comparative experiment shows that the comparison with the model with similar parameters has an advantage,and the subsequent ablation experiments can also see that the performance has been improved to a certain extent when there is no obvious change in the parameters of the relative baseline model.