Self-supervised image denoising and restoration method for atomic force microscopy

Abstract

Scanning probe microscopy (SPM) distinguishes itself from light and electron microscopy by sensing surface interactions with a nanoscale probe, rather than relying on the detection of particles or waves; and has evolved into a versatile tool across several fundamental and applied research fields. SPM uses raster-scanning for image formation, comprising trace (left-to-right) and retrace (right-to-left) scans - this spatial redundancy is however usually not fully taken advantage of. Here, we introduce a self-supervised deep learning framework for Scanning Probe microscopy Image DEnoising and Restoration (SPIDER), by utilizing trace and retrace information. SPIDER improves the signal-to-noise ratio up to 3-fold and accelerates imaging speed up to 6-fold, circumventing the need for a large training set and a ground truth. We further demonstrate that SPIDER enables imaging acceleration by utilizing spatial information from the fast-scan axis to reconstruct missing spatial information in the slow-scan axis in a self-supervised manner. The self-supervised reconstruction is competitive with supervised learning methods. We anticipate that SPIDER will improve SPM imaging and inspire further applications. The authors present self-supervised Scanning Probe microscopy Image DEnoising and Restoration (SPIDER), a deep learning framework that denoises scanning probe microscopy images, enabling imaging acceleration and anisotropic resolution restoration.