N. Kampf, S. Gelman, I. Rosenhek-Goldian, M. Patočka, M. Rios, M. Penedo, G. Fantner, A. Beker, S.R. Cohen, I. Azuri
Weizmann Institute of Science,
Israel
Keywords: AFM, image enhancement, super-resolution, deep learning
Summary:
High quality AFM imaging is time-consuming process. With the current emerging deep learning concept, can low-resolution AFM images be improved by such technique? In this study (1) a model sample was imaged under standard ambient scanning conditions. Both traditional image enhancement methods and deep learning models were applied to improve image resolution and were benchmarked in terms of fidelity, quality, and expert evaluation by AFM experts. The deep learning models outperform the traditional methods, yielding superior results. Furthermore, common AFM artifacts such as streaking, present in the high-resolution ground truth images, were only partially attenuated by traditional methods but were eliminated by the deep learning models. This work demonstrates that deep learning models are superior for super-resolution tasks and enable a significant reduction in AFM measurement time, as low-pixel-resolution AFM images can be enhanced in both resolution and fidelity through deep learning. (1) Beilstein J. Nanotechnol. 2025, 16, 1129–1140.