A Dual Self-Attention Cascaded Adversarial Network For Image Super-Resolution Reconstruction

As a carrier of visual information,image is one of the important sources of information.The quality of the image determines the accuracy of the information and the amount of information,therefore how to improve the image resolution becomes an important part in image processing technology.As a classical problem in computer vision,image superresolution aims at obtaining a high-resolution image from a single low-resolution image.Image super-resolution reconstruction is a research hotspot in the field of image processing.It has been widely used in video surveillance,image compression,and criminal investigation.In recent years,due to the rapid development of deep learning,image superresolution reconstruction method based on deep learning has gradually became a mainstream.On the basis of previous research work,the existing image super-resolution reconstruction methods are introduced and summarized,the shortcomings of the existing image superresolution reconstruction methods are analyzed,and a dual self-attention cascaded generative adversarial network for image super-resolution reconstruction is proposed.A series of experiments are conducted on benchmark datasets to evaluate the validity of the proposed method.Firstly,the main methods of image super-resolution reconstruction are summarized,and three existing image super-resolution reconstruction methods are discussed in detail in this paper.These methods include traditional image super-resolution reconstruction,image super-resolution reconstruction based on convolution neural network and image superresolution reconstruction based on generative adversarial network.The advantages and disadvantages of these methods are also pointed out.Secondly,aiming to address the problems of existing image super-resolution reconstruction methods,the following improvements are made in this paper.(1)The existing image super-resolution reconstruction methods upsample the low-resolution image with a simple interpolation method,or upsample only at the end of the network,typically using a sub-pixel convolution layer or transposed convolution layer to recover the high-resolution result.For large upsampling factors,these methods have a large memory footprint and a high computational cost as it operates on upsampled images,the reconstructed image is more prone to checkerboard artifacts.A cascaded generative adversarial network is proposed in this paper.The generator allows end-to-end refinement for fine-grained high resolution image reconstruction.We use a multi-scale discriminator.The global discriminator focuses on global image structures,which is used to instruct the generator to generate the globally consistent image.The Local discriminator focuses on local image details,which is used to guide the generator to generate rich and meaningful details.(2)The existing image super-resolution reconstruction methods seldom take into account the global dependence between features.As a result,when we super-resolve at a large upscaling factor,the reconstructed image tend to be too smooth that lack high-frequency textures and do not look natural.A self-attention mechanism is introduced and helps with capturing visual features dependencies in the spatial and channel dimensions respectively.The position selfattention module selectively aggregates the features at each position by a weighted sum of the features at all positions.Similar features would be related to each other regardless of their distances.Meanwhile,the channel self-attention module selectively emphasizes interdependent channel maps by integrating associated features among all channel maps.We sum the outputs of the two self-attention modules to further improve feature representation which contributes to more precise reconstruction results.Finally,the proposed method are implemented with Python in Windows 7.The proposed method for image super-resolution reconstruction is evaluated on five benchmark datasets,and show competitive results as compared with 4 traditional image super-resolution reconstruction methods and 13 image super-resolution reconstruction methods based on deep learning.Experiments are conducted to validate the effectiveness,and the proposed method can reconstruct more real and meaningful image details.