Research On Image Super-Resolution Algorithm Based On Generative Adversarial Networks

Image super-resolution reconstruction task is an important research topic in the field of image processing.The main task is to reconstruct high-resolution images with more details from lowresolution images.Although the quality of image super-resolution reconstruction has been greatly improved,it still has research value for improving the visual perception of reconstructed images.In this paper,in order to overcome the fuzzy problem of details in the reconstruction results of traditional super-resolution models,this paper combines the attention mechanism and the idea of contrastive learning to construct a new super-resolution model.First,this thesis builds an asymmetric branch network to extract the features of the input image.By comparing the difference of the feature maps between the two networks,the loss is calculated and the weight of the network is updated inversely,so as to enhance the ability of the network to learn the distribution of image data.At the same time,this thesis designs a feature enhancement module based on the attention mechanism to enhance the intermediate feature maps generated by the asymmetric branch network.The comparison of the final evaluation index scores and the visualization results both prove that the proposed method can effectively improve the quality of image details.Most super-resolution methods are trained on datasets where high-resolution images and corresponding low-resolution images are obtained by a fixed degradation method.However,these methods trained on external datasets will not be able to recover the texture information of the test images if the specific texture information of the test images cannot be found in the dataset.Therefore,combined with the characteristics of the generative adversarial network model,this paper proposes a momentum feature contrast network to generate super-resolution images with rich texture information.The feature contrast network has a momentum-updated siamese structure,which ensures the continuous consistency of the intermediate feature maps while expanding the low-resolution image information.Furthermore,this thesis designs a feature fusion module to process the feature maps of the two branch networks,preserving the global distribution of features and enhancing local highfrequency details.The method proposed in this thesis does not involve high-resolution images in the training process,and belongs to a completely unsupervised super-resolution method.The final qualitative and quantitative experimental results show that the method proposed in this paper has a good improvement in reconstruction quality.In order to further explore the feature map information generated by the feature comparison network and improve the quality of the final image reconstruction,this thesis proposes an orthogonal vector enhancement module.By augmenting the channel information and spatial information of the feature map,the reconstructed basis vectors of the features are learned and mapped into the signal subspace using an orthogonal transformation.Finally,the module enhances the local texture information of the reconstructed image and suppresses the global noise of the image.The experimental results show that the orthogonal vector enhancement module proposed in this thesis effectively improves the texture quality of the reconstructed image,and also has obvious effects on suppressing image noise.