Statistical Latent Manifold-Guided Framework for Generative Super-Resolution

Single-image super-resolution (SISR) has demonstrated remarkable development in computer vision, including medical domain. However, fundamental challenges remain in recovering fine structural details and maintaining anatomical consistency, primarily due to the inherent information limitations of single-view inputs. Data scarcity is a significant challenge in the medical field, where images are often acquired from patients with diverse and complex conditions. Despite the outstanding performance of existing multi-view SR approaches, they typically demand multiple image acquisitions, leading to increased patient exposure and scanning time, an impractical solution for clinical settings. To alleviate limitations, we introduce a novel framework that leverages statistical similarity metrics to guide generative model-based view synthesis from single-view inputs. Our framework features a statistical similarity-aware generative model designed to produce synthesized views that retain local and global consistency with the original image while introducing complementary structural details. These synthesized views are integrated into a multi-view SR pipeline using a custom fusion mechanism that assigns weights to views based on statistical reliability. Experimental results on diverse computer vision datasets demonstrate that our approach significantly outperforms state-of-the-art SISR methods, achieving up to 7.66% improvement in structural preservation and 15.3% reduction in reconstruction artifacts. Our research highlights the effectiveness of SR technology in computer vision as well as the medical domain, demonstrating superior performance even with a limited single image. Further, our advancement holds promise to enhance the accuracy of medical image analysis with a substantial impact on healthcare applications by improving diagnostic capabilities.