A novel single image super-resolution reconstruction model based on edge-enhanced Siamese generative adversarial networks

Abstract

Abstract Currently, most image super-resolution (SR) reconstruction algorithms are supervised, improving the overall visual quality of images based on a large amount of paired low-resolution (LR) and high-resolution (HR) image. However, collecting a large number of paired datasets is challenging, and the quality of the dataset can influence the reconstruction results. Although unsupervised methods have gained significant performance based on synthetic datasets, their effectiveness is relatively poor when applied to real-world or natural data. Focusing on those aforementioned issues, a novel image SR reconstruction algorithm of Edge-enhanced Siamese Generative Adversarial Network (EeSiGAN) is proposed. EeSiGAN belongs to the unsupervised category and does not require the involvement of HR images. Firstly, employing the Siamese GAN as the overall framework ensures the continuous consistency of image feature maps through the collaborative effect of support and main networks. Secondly, fusing the results of the two branches by using the Multi- scale Channel Attention Feature Fusion (MsCAFF) module. In addition, an Edge- enhanced Feature Distillation Block (EFDB) is introduced to enhance edge information in images and optimize the capability in feature extraction of network. Finally, Directional Variance Attention (DiVA) mechanism is used to obtain long-range spatial dependencies and simultaneously utilize inter-channel dependencies to achieve a more discriminative representation, thereby enhancing the overall quality of the recovered images. Extensive experimental results on synthetic and real datasets demonstrate that compared to other state-of-the-art unsupervised SR networks, the proposed EeSiGAN algorithm exhibits better performance in reconstructing images, producing clearer details and textures in the enlarged images.