High-resolution transmission electron microscopy (HRTEM), with its outstanding spatial and temporal resolution, has enabled unprecedented understanding of the atomic structures of materials and how structure relates to properties and functions. However, capturing dynamic processes with high temporal resolution inevitably leads to severely degraded HRTEM images due to experimental conditions and imaging parameters, which substantially limit accurate structural analysis. In this paper, we develop a structure-preserving HRTEM restoration framework that enhances low-quality HRTEM images with blurred or incomplete atomic arrangements using a generative deep learning approach while maintaining physical fidelity. Specifically, we propose HRTEM-GAN, a cycle-consistent generative framework that operates under unpaired training conditions and performs patch-level distribution modeling between high- and low-quality image domains, while explicitly incorporating frequency-domain constraints to preserve atomic-scale structural fidelity. This design enables effective restoration of low-quality HRTEM images, yielding structurally coherent atomic lattices that are fully suitable for subsequent recognition and quantitative analysis. The proposed method is validated on a real experimental dataset acquired during in situ imaging of Au catalysts under CO oxidation conditions. Compared with representative methods, HRTEM-GAN achieves substantial improvements in image restoration quality and consistently enhances downstream atomic column recognition performance. These results demonstrate the potential of the proposed framework to facilitate reliable atomic-scale analysis in HRTEM studies.