Research On Image Super-resolution Based On Deep Learning

With the development of the information age,digital devices are becoming more and more popular,and the role of digital images in information dissemination is becoming more and more important.There is an increasing demand to improve the quality of digital images.Image super-resolution reconstruction aims to recover high-resolution images of the same scene using one or multiple low-resolution images combined with prior knowledge.Deep learning aims to automatically learn the relationship between input and output directly from the data,which overcomes the problem of artificial design features.A large number of empirically proven image super-resolution reconstruction algorithms based on deep learning have achieved promising results than traditional counterparts.Therefore,super-resolution reconstruction based on deep learning has become a research hotspot.After the convolutional neural network was introduced into the field of super-resolution reconstruction,a large number of improved algorithms appeared.The algorithm based on residual learning and attention mechanism showed excellent performance in improving model performance and accelerating network training process.Therefore,this paper focuses on the above two mechanisms.Firstly,the classical convolutional neural network structure is discussed,and the analysis of its insufficiency is carried out,and the residual learning mechanism is introduced.Focusing on residual learning,from the deep residual network to the Inception structure,the advantages and disadvantages of the two are analyzed,and the improved dense residual structure is proposed which uses the same topology for different branches of the Inception network and adopt dense connections within the branch to improve the generalization ability and scalability of the network.In the upsampling phase,the sub-pixel convolutional layer is used to improve the influence of the checkerboard artifacts and improve the reconstructed image quality.Then,according to the attention mechanism,three classic attention models are introduced,and the expectation maximization mechanism is introduced for the problem of large amount of attention map calculation.On this basis,the local expectation maximization mechanism which is more suitable for super-resolution reconstruction tasks is proposed.The image is divided into different regions by edge detection,a set of compact bases is iterated in different regions,and then the attention mechanism is operated on this group,which greatly reduces computational complexity.Combining the enhenced dense residual structure with the local expectation maximization attention mechanism,the local expectation maximization attention model(LEMA)is proposed.An improved perceptual loss function is proposed,combined with an MS-SSIM loss function that reflects the image structure,in order to better recover the high-frequency details of the image.The experimental results show that the model can restore more realistic and natural texture details,and the reconstructed image subjective quality is significantly improved.