| Super resolution reconstruction aims to restore a low-resolution image to a high-resolution image,and retain high-frequency details in the low-resolution image as much as possible.As a basic low-level problem in computer vision,super-resolution has received extensive attention in recent years,and has important applications in image compression,medical imaging,remote sensing imaging and so on.With the rapid development of deep learning technology,the superresolution method based on deep neural network has made many achievements.However,super-resolution still faces some problems in practical applications.On the one hand,natural images often contain unknown degradation.When the degradation mode of the input image does not match the trained model,the performance of the algorithm will be seriously degraded.On the other hand,with the deepening of the network structure,the amount of model parameters and computation becomes more and more huge.How to improve the computational efficiency is the focus.Therefore,we conduct an in-depth study on these issues.The main contents include:1)A super-resolution reconstruction method based on contrast learning estimation of image degradation is proposed to solve the problem of blind super-resolution of input image with unknown degradation.In order to deal with different kinds of degraded images,we construct a blind super division network which includes degraded estimation branch and conditional reconstruction branch.First,the degenerate estimation branch makes use of the characteristics of the contrast learning discriminant method to implicitly estimate the degenerate representation,so that fuzziness and noise can be treated in the same frame.At the same time,using the pretraining strategy of extended data sets can better play its advantages of unsupervised learning and improve the ability of feature extraction.Secondly,the conditional reconstruction branch uses the estimated image degradation representation to dynamically adjust the characteristics of each layer of the network through the degraded channel attention module and the degraded spatial attention module,so as to effectively adapt to different degraded input images.2)Aiming at the lightweight problem of super resolution network,we propose a new lightweight super resolution algorithm based on efficient feature distillation network.Among them,the proposed efficient feature distillation module further refines the features while retaining the useful features by flexibly segmenting and connecting the feature maps,reducing the redundancy of network parameters.The designed enhanced pixel attention module can generate effective pixel level attention weight by squeezing and stimulating the feature map channel,adjust parameters adaptively,and enhance the network expression ability.In addition,we explore the impact of different activation functions on the network,and proposes a multiscale hot start training strategy.Through multiple training of the network with image blocks of different sizes,the reconstruction quality can be further improved without changing the computational efficiency.3)We performed a practical deployment and application of the proposed super-resolution model.First,the Py Torch model is exported to the open neural network exchange format ONNX as an intermediate model for cross-platform model deployment.Second,a localized model inference engine is built from the ONNX model using the Tensor RT framework,and GPUaccelerated inference and optimization are performed for specific hardware device platforms to achieve high-performance model inference.In addition,we designed a graphical superresolution system tool,which can select different data and models for super-resolution testing and application through intuitive parameter settings,and built a simple web application using Streamlit framework to provide users with a more convenient super-resolution reconstruction application. |
