Study On Image Super-resolution Reconstruction Algorithm Based On Attention Mechanism

Image super-resolution is a classic problem in computer vision.Thanks to its unique and outstanding image restoration capabilities,it is widely used in a significant number of fields.With the increasing image super-resolution research based on deep learning,image pixel restoration has been significantly improved.However,the existing algorithms often lack indepth mining of image features,so the effect of image restoration still has some shortcomings.In response to the above problems,this paper starts with the characteristics of the image,introduces a variety of attention mechanisms,and improves the image super-resolution algorithm based on the generative adversarial network.The main work of this paper is as follows:1.The paper improves the SRGAN model and proposes a method for super-resolution reconstruction of remote sensing images based on edge loss.The method is based on the generative adversarial network.In addition to considering the color and perception characteristics of the image during feature extraction,the method also pays more attention to the remote sensing image’s edge information,and feeds back the edge information to the network in the form of an edge loss function.Through experiments on a large number of remote sensing image datasets,it is proved that the remote sensing image super-resolution method based on edge loss can effectively improve the quality of remote sensing image resolution restoration.2.The paper improves the SRGAN model and proposes a saliency-guided image superresolution reconstruction method.To achieve the purpose of improving the image resolution,the method detects the saliency area of the image through the deep network model and designs the saliency loss of the network to pay more attention to the saliency features of the image.The experimental results on the saliency detection databases,the image super-resolution standard databases,and the remote sensing image databases prove the method’s superior performance.3.An enhanced full-resolution residual network applied to image super-resolution is proposed.The network structure is inspired by the full-resolution residual network of semantic segmentation.The network structure is modified and improved to design a generator network suitable for image resolution.Besides,the self-attention mechanism is introduced to make the network automatically assign different weights to channels and spatial according to feature characteristics to improve feature extraction quality.The experimental results on the image super-resolution standard datasets prove the superiority of the algorithm.