Research On Image Super-Resolution Based On Convolution And Transformer Feature Aggregation Network

Image super-resolution technique is a method to recover high-resolution images from low-resolution images,and it can improve the visual effect of images.With the development of technology,images have become an important vehicle for acquiring information.However,due to the complex acquisition environment,insufficient hardware performance or improper operation,the acquired images usually have low resolution,blurred details and missing key information,which cannot achieve the ideal visual effect.At this point,image super-resolution techniques come into play.At present,the method based on convolutional neural network is the mainstream image super-resolution method,but it has some defects,such as high model complexity,small perceptual field,and lack of adaptability.In order to solve these problems,this paper proposes a network model based on the combination of convolution and Transformer to study and improve the image super-resolution algorithm.The main works are:(1)A feature aggregation network CTCA is designed.The model fuses the convolutional branch and Transformer branch together,where the convolutional branch is responsible for extracting local features and detailed textures,and the Transformer branch is responsible for capturing long-distance dependencies and weak textures.Information interaction and discrepancy amplification between the two branches is performed by a channel attention fusion module.This module can amplify the unique characteristics of each of the two features by modeling the weighting relationships between the channels while preserving the common action of both features.A richer and more balanced information representation is achieved by fusing the features of both mechanisms in a multi-level manner.(2)A Transformer module is improved,which alternately uses the global multi-headed self-attention mechanism and the local windowed multi-headed self-attention mechanism to extract image features.The global self-attention mechanism can exploit the similarity textures existing inside the image to enhance the global information and can capture long-range dependencies,while the local windowed self-attention mechanism can exploit these similarity textures to mine more details.This can extract rich global and local information from image features simultaneously.(3)A convolution module is improved to extract global features in an image by setting a deep convolution with a large kernel in the spatial attention layer,and using dynamic weights to determine the importance of different features at the spatial level and find similarity textures.This is combined with a proposed convolutional layer structure similar to the feed-forward network in Transformer to further refine the previously extracted important information and textures in order to recover more high frequency features in the image.By using both modules interchangeably,rich global and local features are extracted.