| With the development of technology and the times,there is an increasing demand for high-definition images in many AI-based technological tasks.Obtaining high-definition images through hardware is too expensive,so super-resolution technology reconstructs high-resolution images from low-resolution ones at the software level.Deep learning has strong learning ability and portability,and since its application in super-resolution reconstruction,research based on deep learning for super-resolution reconstruction has gradually become the mainstream in the field.Although super-resolution reconstruction technology has made good breakthroughs in recent years,there are still deficiencies in the quality of reconstructed images,such as artifacts,structural distortion,and loss of details.In order to solve these problems and reconstruct better quality images,this paper introduces the coordinate attention mechanism,dense connection,and multi-scale technology into SRGAN and reconstructs a super-resolution model based on the generative adversarial network.This paper starts from the theoretical research of super-resolution reconstruction,based on the development and research status of super-resolution,and mainly does the following two works:(1)Aiming at the problem that the parameters of existing image super-resolution reconstruction models are too large to be applied in reality,a single-image super-resolution reconstruction model CSRGAN(generative adversarial network based on coordinate attention mechanism)is proposed.By optimizing the generator of SRGAN,the coordinate attention mechanism is combined with the residual network to construct the CR module,which promotes the information flow between channels and enhances the feature selection ability of the network.The main network is constructed with a hierarchical feature fusion structure to improve the utilization of early features in deep networks,and a large number of long-short skip connections are used to alleviate gradient disappearance and improve the network convergence speed.The model is tested on Set5,Set14,BSD100,and Urban100 datasets and compared with RFB-ESRGAN,ESRGAN,and other models.The peak signalto-noise ratio(PSNR)and structural similarity(SSIM)are improved,and the model parameters are greatly reduced.The reconstructed images are clear,structured,and rich in detail.(2)Aiming at the problem of insufficient image clarity and missing texture details in reconstructed images,a super-resolution model based on multi-scale attention mechanism(MFCSRGAN)is proposed.A multi-scale attention module(MFCR)is constructed by combining residual network,multi-scale feature mechanism,CBAM attention mechanism,and dense connection.By improving the residual block,feature expression is enhanced,and more pixel information is retained.Different sizes of convolution kernels are used to obtain feature information under different sizes of receptive fields to increase feature diversity.The basic module CR module in CSRGAN is optimized,and a spatial attention module is added after the coordinate attention module to enhance spatial learning.The model is tested on multiple standard datasets and compared with CSRGAN,ESRGAN,and other models.The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)are improved,and the reconstructed images are further improved in terms of clarity and structural integrity. |
