Research On Image Super Resolution Algorithm Based On Deep Learning

In image processing tasks,single image super-resolution is one of the most important technologies.Now,deep learning-based methods show advantages in various fields,but there are still some problems in image super-resolution reconstruction such as poor reconstruction effect,difficulty in network training,and large memory usage.This thesis starts the research from the image reconstruction quality,model lightweight and image up-sampling method.The main work is as follows:(1)In this thesis,to solve the troubles that deepening of the network layer easily leads to a large amount of calculation and it is difficult to deploy on mobile devices.So we pose an algorithm based on the Dense Residual Attention Network(DRAN).The main body of DRAN adopts a residual neural network,which establishes the connection between the current layer and all the previous layers through dense connection and can achieve feature reuse,relieve the slow convergence during training.In addition,the channel attention mechanism can model the correlation between feature channels,so that the network focuses on useful information,ignoring useless information,and improving the accuracy of image super-resolution.The test results prove that DRAN has strong robustness,less calculation,and the reconstructed image has a clear visual effect and is suitable for mobile devices.(2)Many image super-resolution means mostly train the network model separately for each factor,which not only complicates the network training process,but also easily causes waste of resources,which does not meet the actual use scenario requirements.To this end,this thesis proposes a multi-factor image super-resolution network model(MFN),which takes the magnification factor as input,dynamically predicts the weights of the upsampling filter,and uses these weights to generate high-resolution images of the corresponding magnification,which is suitable for image super-resolution reconstruction under arbitrary integer or non-integer situations has strong flexibility.In addition,innovating based on the basis of the dense residual attention network,the information distillation network(IDN)is used as the feature extraction module to gradually extract multi-scale spatial features,and feature fusion according to feature importance,which significantly improves visual quality and reduces the amount of calculation.Finally,this thesis improves the channel attention mechanism,adds contrast perception factors,and enhances detailed information such as image structure and texture.In all,this thesis proposes a multi-factor image super-resolution network based on information distillation—IDMF-SR.The experimental results show that on the Set14 dataset,the PSNR value of IDMF-SR reaches 25.50 d B,which is an improvement of 0.07 d B compared to DRAN,which effectively refines the image details;IDMF-SR reduces 69.8% of the parameter amount compared with Meta-SR.The PSNR value increased by 2%.