Research On Image Super-resolution Algorithm Combining Attention Mechanism And Its Application

Image super-resolution(SR)is a technology to process low-resolution(LR)images into high-resolution(HR)images through computer technology,and restore some details in the images.The influence of super-resolution algorithm in the human world is very profound,and its application scenarios are very wide,including satellite imaging,medical image analysis,video surveillance,autonomous driving and many other fields.At the same time,with the continuous progress of human science and technology,and more attention has been paid to the quality of images,image super-resolution technology has also been developed.However,for any low-resolution image,there are countless high-resolution images corresponding to it.The early algorithm mainly focuses on the interpolation method,which has high execution efficiency and simple algorithm implementation,but also has a big bottleneck.As a result,more and more people are starting to look into learning-based method.At the same time,the development of deep learning has brought more possibilities for the application of super-resolution.The main contents of the dissertation are as follows:(1)The channel attention mechanism mainly used in the low-level vision task ignores the information at the spatial level,which will limit the range of adaptive selection of the network in the process of training.At the same time,it is found that most of the loss functions used in network training pay the same attention to high frequency information and low frequency information,which hinders the network's learning of high frequency information.Therefore,a new method is proposed.Region-level channel attention(RCA)mechanism,which enables the network to allocate different attention to channels in different spatial regions.At the same time,a high-frequency aware loss is proposed to enhance the attention of the network to the high frequency details.Many experiments are set up to study the reliability of the proposed algorithm.At the same time,the images generated by this method have better visual effect compared with other contrast algorithms,and the objective indicators are also better than other contrast algorithms.(2)Generative adversarial networks have been widely used in the field of super-resolution.Although the results generated by the super-resolution model obtained by using adversarial training can be greatly improved in vision,there is usually a big difference with the original image in semantics.Therefore,a structure preserving loss is designed on the loss function to make the network pay more attention to its structure information in the process of training.At the same time,a wide channel activation structure is introduced into the generator so that the information in the shallow layer of the network can be well transmitted to the deep layer of the subsequent network.Perceptual loss is also used during training to make the generated results more detailed.A number of test sets are tested,the result of the algorithm can not only ensure the visual effect of the image,but also alleviate the distortion.