Research On Super-resolution Reconstruction Based On Deep Learning

As people’s requirements for high-quality images become increasingly urgent,people are increasingly concerned about how to obtain larger resolution and higherquality images without updating hardware devices.As a popular low-level computer vision task,super-resolution reconstruction is applied to all aspects of our lives,such as video surveillance,satellite remote sensing,etc.In recent years,the development of deep learning has made the research on super-resolution reconstruction achieved remarkable results.However,super-resolution reconstruction still faces many challenges.The super-resolution reconstruction algorithm based on deep learning has problems such as poor generalization ability and large amount of calculation,which greatly limits the application of super-resolution reconstruction in real scenes.The main contributions of this article are as follows1.In this paper,we first analyze the existing shortcomings of the current mainstream deep learning-based super-resolution algorithms under real scene conditions.These shortcomings are mainly because the model has too much calculation and cannot be directly applied to real scenes.In view of the large amount of calculation in deep neural networks,we introduce a weighted residual network structure and multiple feature fusion modules,and design a lightweight super-resolution network based on feature fusion.The designed super-resolution network not only requires less calculation,but also has a performance comparable to state-of-the-art methods.2.In order to enable the deep super-resolution network to be applied to real scenes,super-resolution reconstruction can be directly performed on photos taken by smart cameras and other handheld devices.We introduce a kernel estimation method based on Generative Adversarial Network,and combine it with the super-resolution network proposed on the appeal.The data generated by the kernel estimation method can be used for the training of the appeal model,and the trained model can be directly applied to the real image.3.In order to further reduce the amount of network calculations,we also propose a knowledge distillation method based on feature similarity.By decomposing or binary quantization of the convolution operation in the deep neural network,we can further compress the proposed lightweight model.The proposed knowledge distillation method based on feature similarity can effectively reduce the accuracy loss caused by model compression.