Research On Single Image Super Resolution Based On Deep Learning

With the rapid development of the Internet and multimedia technologies,visual data including images and videos have rapidly increased.The corresponding data processing algorithms are also becoming more and more important.Massive visual data often contains a wealth of useful information,and processing them through computer vision helps to better serve humans.The goal of super resolution is to improve the visual quality by increasing the resolution of visual data,which is beneficial to computer vision to better capture the useful information.In this paper,based on the deep learning algorithm,the single image based super resolution is deeply studied.The main work and the corresponding conclusions are summarized as follows:1.For the problem of objective function,this paper discusses the impact of different loss functions on the training of super resolution networks.Although the mean square error function will cause the super resolution model to output an image with a high PSNR value,the texture details tend to be blurred and the visual effect is poor.The loss calculation method including perceptual loss and texture loss are to obtain the consistency of the prediction image and the ground truth image from the feature level,so that a better visual effect will be obtained.The total variation loss function preserves the edges of the image while removing noise.Based on the generative adversarial network,the generated image is discriminated in two dimensions of image and feature.After the two networks reach a Nash Equilibrium,the generator can learn the distribution close to the real data.Finally,the paper also discusses the impact of different loss function combinations on the training effect,which promotes the model to achieve better super resolution effects.2.For the problem that high-frequency texture information in images is difficult to recover,this paper proposes a method of embedding attention mechanism to enhance high-frequency information recovery.The attention mechanism is to make a decision at the pixel level through a fully convolutional network to determine whether each pixel is in a high frequency region.After attention mechanism embedding,the corresponding gain and weakening are performed,beneficial for the recovery of high-frequency information.This paper also tried a variety of algorithms to improve quality of the attention generation network to further improve performance.The corresponding experiments show the difference in high-frequency detail recovery before and after the attention mechanism embedding,and the PSNR and SSIM are also improved on different datasets,thus demonstrating the effectiveness of the attention mechanism.Finally,the performance of the entire network framework is demonstrated by comparison with other algorithm models.3.For the problem that the super resolution network is slow to train and difficult to converge to a better local minimum value,the curriculum learning is proposed to train the super resolution network.Curriculum learning mainly involves curriculum setting and corresponding curriculum learning methods.In the aspect of curriculum setting,the image texture features are extracted based on the gray level co-occurrence matrix,and then the clustering method based on density peak is used to realize the division of the training set.As for the curriculum learning method,the probability distribution of the sample is dynamically allocated based on the slope of the learning curve on different training subsets.Experiments show that the gray level co-occurrence matrix and the clustering method based on density peak can realize a better division of the training set on the texture complexity level.The improved curriculum learning method proposed in this paper can accelerate the convergence of the model while avoiding the forgetting of the ‘knowledge' and improving the overall performance of the model.