Presentation Information
[9a-E310-9]Evaluation of the generalization capability of real-time SPM image denoising AI
〇ZHUO DIAO1, Taisei Noji1, Masahiro Ohara1, Masayuki Abe1 (1.Osaka Univ.)
Keywords:
scanning probe microscopy,denoising,neural network
Real-time denoising of scanning probe microscopy (SPM) images is essential for improving the reliability of AI-driven autonomous experiments. However, the absence of ground-truth experimental images limits the applicability of supervised learning, while simulation-based training may mistakenly remove unknown surface structures as noise. In this work, we propose a real-time SPM denoising (RTSPM Denoise) framework that requires only experimental images for training. Two U-Net-based pipelines are investigated: (1) Recorrupted-to-Recorrupted (R2R) learning using artificially re-corrupted experimental images, and (2) Noise2Noise (N2N) learning using pseudo-clean images generated by the pretrained SwinIR model as supervision. The two approaches are systematically compared in terms of denoising performance and structure preservation. Furthermore, a large experimental SPM image dataset covering diverse material surfaces is constructed to evaluate generalization under out-of-distribution (OOD) conditions. The proposed framework provides a practical solution for robust real-time denoising in autonomous SPM systems.