Research On Image Super-resolution Reconstruction Method Based On Transforme

With the rapid development of science and technology,human beings have entered a new era of information technology,and images,as one of the important carriers of information storage and transmission,are the main source of information acquisition and exchange for human beings.High-resolution images mean richer external information,however,in real life,images can be affected by complex and diverse factors leading to image quality degradation,and the integrity and accuracy of information transmission cannot be guaranteed.Therefore,improving image resolution through super-resolution reconstruction techniques has become a hot issue for research in the field of computer vision.Image super-resolution reconstruction aims to reconstruct a high-resolution image containing clear detailed features from a given low-resolution image.In recent years,with the development of deep learning techniques,convolutional neural networks have achieved remarkable success in image super-resolution reconstruction tasks due to their excellent feature extraction ability.However,due to the limited perceptual field of convolutional neural networks,existing deep learning-based super-resolution reconstruction methods tend to expand the perceptual field by stacking a large number of convolutional layers to obtain global information,which not only brings problems such as gradient disappearance and increase in the number of parameters,but also gradually exhausts the ability of the network model to obtain information as the number of layers deepens.The successful application of Transformer in various computer vision tasks proves its powerful characterization capability in the image processing field as well.In view of this,this paper introduces Transformer to image super-resolution reconstruction tasks and proposes two Transformer-based methods for image super-resolution reconstruction,with the following main research work:(1)A Vision Transformer based image super-resolution reconstruction method is proposed.The Transformer module is used to replace the convolutional neural networks in the traditional deep learning method to extract the features of the image,and the self-attention mechanism is used to capture the long-space dependence relationship between features in different locations of images,so as to maintain the spatial structure of reconstructed images.At the same time,the convolution layer is used to extract the features of the image,and the global residual connection is used to fuse the low frequency features and high frequency features,which improves the reconstruction performance.The experimental analysis on multiple benchmark data sets shows that the proposed algorithm has better performance than the traditional convolutional neural network reconstruction algorithm,which proves the effectiveness and practicability of the network.(2)An image super-resolution reconstruction method combining lightweight neural network and Transformer is proposed.By using depthwise separable convolution instead of ordinary convolution and combining the Swin Transformer module,the network can achieve excellent reconstruction performance with a small training data set,reduce the number of model parameters and computational complexity,and improve the reconstruction efficiency.Secondly,the global residual and local residual are combined to enhance the feature fusion of different levels,so that the model converges faster and is easier to train.The experimental analysis on a large number of benchmark datasets shows that the image super-resolution reconstruction algorithm proposed in this chapter significantly improves the reconstruction performance,and achieves the best balance between the amount of parameters and the reconstruction performance.