Research On Image Super-Resolution Reconstruction Based On Deep Learning

Image is the most commonly used information carrier in human social activities.It can intuitively and vividly describe the information of objective objects.Generally,high-resolution images present more detailed information,so scholars have proposed a super-resolution reconstruction technique,which can reconstruct low-resolution images into visually pleasing high-resolution images through certain algorithms.The technology is widely used in many fields such as medical imaging,satellite imaging and security imaging,etc.The super-resolution reconstruction model based on the deep network structure can better learn the inherent rules and hierarchical expression of the data,which is helpful for the recovery of high-frequency image details.However,the existing deep learning model has the problems of complicated network structure,large amount of calculation,and treating the low-frequency component and high-frequency component of the feature map equally during image reconstruction,which makes it difficult to better reconstruct the image details.In order to solve these problems,this paper designs a super-resolution reconstruction model based on deep learning,which is studied from the aspects of network structure design and improvement of loss function.The main work of this article includes the following aspects:1.In order to better learn to use the high-frequency information of the image,this paper introduces the feature map attention mechanism,which can use the interdependence between the feature channels to adaptively adjust the channel features to achieve the purpose of enhancing the ability to express features,thereby recover more contours,textures and other details.2.In order to make better use of different levels of feature information,an information distillation network that includes an information enhancement units and compression units is designed,with the combination of local shallow network features and local deep network features used to obtain more effective information and packet convolution used to increase the training speed of the network.In the case of using fewer convolutional layers,the quality of image reconstruction can also be guaranteed.3.Sub-pixel convolution in the reconstruction layer of the network is introduced,so that the network can directly perform feature extraction operations on the input low-resolution image,avoiding the interpolation pre-processing of the input image by the network and reducing the amount of calculation.4.The convergence characteristics of two pixel-by-pixel loss functions MAE and MSE that are common in image super-resolution reconstruction networks are studied,combining the advantages of the two loss functions to train the network,which makes the network play better performance.The experimental results on Set5,BSD 100,Urban100 and Manga109 datasets show that the network proposed in this paper can effectively improve the visual effect of images,and also prove the role of the feature map attention mechanism in the image super-resolution detail reconstruction.