CSM-SR: Conditional Structure-Informed Multi-Scale GAN for Scientific Image Super-Resolution

Abstract

High-resolution image reconstruction is essential in scientific imaging, where preserving structural details directly affects downstream analysis. While Generative Adversarial Networks (GANs) have significantly improved single image super-resolution (SISR), existing approaches often fail to preserve fine structural information in domain-specific data such as scanning electron microscopy (SEM) images. To address this, we propose CSM-SR, a Conditional Structure-Informed Multi-Scale Super-Resolution GAN with three key components. First, it employs an encoder-conditioned feature injection mechanism using a pre-trained EfficientNet-B4 to extract global structural priors directly from low-resolution inputs, removing the need for high-resolution reference images during inference. Second, it introduces a hybrid Semantic Structural Loss (SSL), combining perceptual, contextual, gradient, second-order gradient, structural similarity, texture, and colour consistency terms, with fixed weights selected through grid search on a held-out development set. Third, it incorporates a progressively growing Multi-Scale PatchGAN discriminator to enforce both local texture realism and global structural consistency across multiple resolutions. Trained on a public SEM nanoscience dataset and evaluated on SEM data plus five standard SR benchmarks, CSM-SR achieves state-of-the-art PSNR, SSIM, and LPIPS, with improvements of up to 1.42 dB PSNR and 3.1% SSIM over the strongest baseline.