Image Super-Resolution Reconstruction Based On Generative Adversarial Networks

As an efficient and fast information transfer medium,image is of great importance for people to obtain and transfer information.Due to the inadequate hardware conditions of the image acquisition equipment itself,and the interference of human factors and environmental factors,the acquired image is often a low-resolution image obtained from the original image through a series of degradation processes,which is prone to distortion,blur,noise and other phenomena.The image super-resolution reconstruction technology based on deep learning can improve the resolution of images by directly extracting the features of low-resolution images through algorithmic modeling without relying on the original image acquisition equipment,which is an image processing technology with high cost performance and high feasibility.In this paper,based on the research on the theory related to generative adversarial networks and attention mechanism,the degradation operation,network structure and loss function in the super-resolution reconstruction algorithm are optimized to achieve a better reconstruction effect in combination with the practical application scenario of the research group’s personalized ceramic customization platform.The main contents of this paper are as follows:(1)Considering that most current image super-resolution reconstruction algorithms use a fixed bicubic interpolation algorithm to downsample the original image,which leads to the problem of prediction value deviation and performance degradation when the superresolution algorithm deals with more complex degradation scenes in the real world,for this reason,we propose a second-order degradation model,which extends the traditional interpolation algorithm into a two-stage degradation process,and the complexity and practicality of the degradation operation are enhanced by adding random blurring,downsampling,noise and compression operations in each degradation stage,which enables the super-resolution algorithm to learn more complex degradation relationships and further improves the performance and generalization ability of the algorithm.The experimental results show that the super-resolution reconstruction algorithm using the second-order degradation model can handle the degradation processes such as scaling,noise,blur and compression more effectively.(2)To address the shortcomings of generative adversarial networks,which rely too much on convolutional operations and thus cannot fully capture the features of images at each scale,we propose an image super-resolution reconstruction algorithm HA-ESRGAN based on hybrid attention mechanism and generative adversarial network.The algorithm constructs a multi-scale hybrid attention residual group,and integrates spatial attention,channel attention and self-attention mechanisms into the residual structure unit to accurately extract the spatial features,channel features and global features of the image,while reducing the amount of network computation.In addition,a discriminator based on the Unet structure and spectral normalization layer is used to improve the stability of the network during training.The experimental results show that under the condition of 2 times reconstruction,compared with other advanced algorithms,the PSNR and SSIM metrics of the proposed algorithm on four benchmark datasets are increased by 1.65 d B and 3 % on average,and the NIQE is decreased by 12.1 % on average.Under the condition of 4 times reconstruction,the PSNR and SSIM metrics of the proposed algorithm are increased by1.14 d B and 3.7 % on average,and the NIQE is decreased by 15.4 % on average,and the reconstructed image has better visual perception with more accurate texture details,brightness and saturation.(3)In order to meet the practical needs of personalized ceramic customization application that need to improve the clarity of ceramic patterns,we construct an online web super-resolution platform,first analyze and design the functions and architecture of the platform,then deploy the proposed super-resolution model in the platform,and finally combine the main functions of the platform with personalized ceramic customization application to test and visualize the results,and the test results show that the clarity of ceramic patterns uploaded by users have been effectively improved by the platform,and the platform has practical application value.