Research On Lightweight Image Super-resolution Reconstruction Method Based On Attention Mechanism

With the gradual development of the information age,the demand for high image resolution is also becoming stronger.Due to the high cost and bottleneck of the solution to improve the resolution of the captured image by improving the hardware,and people also have the need to increase the resolution of the acquired image,the super-resolution reconstruction technology,a technical solution that improves the image resolution through digital signal processing,has attracted great attention and in-depth research from the academic and commercial circles at home and abroad.Image super-resolution reconstruction method based on deep learning is a type of method that has emerged in recent years.Such algorithms mainly use convolutional neural networks to learn the mapping relationship between low-resolution images and high-resolution images from a large number of low-resolution-high-resolution image pairs,and apply this mapping relationship to the image that needs super-resolution reconstruction,a high-resolution image will be obtained through inference.Compared with previous super-resolution reconstruction methods,this type of method has greatly improved the accuracy and visual effects of reconstructed images.With the prevalent of mobile devices,the demand for super-resolution on devices with limited computing performance and storage space has become stronger.However,some existing deep learning-based image super-resolution reconstruction methods are limited in applicability due to the large amount of model parameters and high operating memory usage.In addition,the attention mechanism module in deep learning can greatly improve model performance under the premise of a small increase in model parameters.However,the design of the existing attention mechanism lacks specificity for super-resolution reconstruction tasks.Therefore,how to design a lightweight image super-resolution reconstruction method based on the attention mechanism requires in-depth exploration and research.In view of the problems lie in the existing algorithms,this thesis uses the ideas based on lightweight convolutional neural networks and the attention mechanism to study the problem of image super-resolution reconstruction.The main works of this thesis are as follows:(1)Focusing on the problem of limited model performance caused by the mismatch of super-resolution reconstruction algorithm model training and test environment when using the channel attention mechanism in existing methods,this thesis proposes a reconstruction strategy based on local feature enhancement,which can enhance the attention of the attention mechanism to the local features of the image and make full use of model performance.In order to solve the problem that the large amount of network model parameters used in the existing convolutional neural network based super-resolution algorithm leads to the limitation of the use scene,an expanded distillation residual network is proposed to achieve super-resolution reconstruction of high-quality images.At the same time,it reduces the amount of parameters and calculations required by the model.(2)Aiming at the existing super-resolution reconstruction algorithms ignore the relationship between the features in the model when using the attention mechanism,this thesis proposes a feature correlation attention mechanism that improves the network's ability to select and distinguish features.This thesis also uses this attention mechanism to construct a lightweight perceptual image super-resolution reconstruction model,which is used to restore low-resolution images into high-resolution images with more realistic visual effects.In addition,this thesis designs a feature block reconstruction strategy to further improve the performance of the model with feature correlation attention mechanism.Sufficient experimental results prove that the models in this thesis have excellent effects under multiple data sets and multiple evaluation indicators.