Unsupervised denoising of STEM images for enhanced atomic quantification

Scanning transmission electron microscopy (STEM) is essential for atomic-scale materials characterization, yet its quantitative accuracy is often limited by shot and scan noise. Here, we apply an unsupervised denoising framework based on the UNIT architecture, trained using unpaired experimental and simulated images. By integrating spatial- and frequency-domain loss functions, this approach overcomes the traditional trade-off between preserving local defect features and global lattice periodicity. Moreover, the denoised images exhibit significantly enhanced performance in atomic localization and elemental classification, enabling robust automated identification of point defects. These results demonstrate the potential of this method for accurate and reliable structural quantification.