Computer vision distortion correction of scanning probe microscopy images

Abstract

Since its inception, scanning probe microscopy (SPM) has established itself as the tool of choice for probing surfaces and functionalities at the nanoscale. Although recent developments in the instrumentation have greatly improved the metrological aspects of SPM, it is still plagued by the drifts and nonlinearities of the piezoelectric actuators underlying the precise nanoscale motion. In this work, we present an innovative computer-vision-based distortion correction algorithm for offline processing of functional SPM measurements, allowing two images to be directly overlaid with minimal error – thus correlating position with time evolution and local functionality. To demonstrate its versatility, the algorithm is applied to two very different systems. First, we show the tracking of polarisation switching in an epitaxial Pb(Zr0.2Ti0.8)O3 thin film during high-speed continuous scanning under applied tip bias. Thanks to the precise time-location-polarisation correlation we can extract the regions of domain nucleation and track the motion of domain walls until the merging of the latter in avalanche-like events. Secondly, the morphology of surface folds and wrinkles in graphene deposited on a PET substrate is probed as a function of applied strain, allowing the relaxation of individual wrinkles to be tracked.

Effect of uniaxial strain on the optical Drude scattering in graphene

“…3(c) for the topographical micrograph at 2% nominal strain, a number of small folds and wrinkles in the graphene appear to have relaxed. To quantify this relaxation behavior, we performed a differential analysis of the topographical images using an in-house developed drift correction algorithm [16], which allows changes from one scan to another of the same area to be determined with sub-nm precision. Fig.…”

Correlative imaging of ferroelectric domain walls

“…The isotropic SHG image is then binarized, as shown in Fig. 5 b, and both images are subjected to a computer vision based image registration algorithm based on the effective correlation coefficient method and a homography matrix transformation 29 , 31 , 39 . The difference between the proxy and SHG images is shown in Fig.…”

“…The key paradigm behind the proposed workflow is that of a data-driven analysis approach. Both SHG and SPM data acquisition produce two-dimensional maps that are spatially referenced—and can therefore be directly compared after an appropriate coordinate map transformation 29 . Once generated, the corrected observable maps can be stacked into a single aggregated dataset, where each spatial point or voxel contains a heterogeneous vector combining the parameter space of both techniques.…”

“…A solution to this challenge is provided by the recent advances in the use of computer vision and machine learning for the investigation of ferroelectric materials 25 – 28 . Computer vision can be used to match data based on common features across the two completely different techniques 29 . Machine learning techniques can not only identify and quantify behaviours in large data sets across multiple parameters, but through additional pre-processing methods such as dimensional stacking can provide physical insights into the intrinsic correlations at multiple length scales and across characterization methods 30 .…”

Sequential piezoresponse force microscopy and the ‘small-data’ problem

“…However, with an everincreasing data generation tendency in scientific instruments, the necessary data organization and analysis techniques have also come a long way. [1][2][3][4][5][6][7] This is also valid for high-throughput combinatorial searches in materials science, as well as chemistry, in which a large number of parallel or complementary properties may be measured with changes in composition or processing variables. Such data is used to identify previously unknown trends or materials-property connections, but a problem exists in efficient analysis and visualization of the data.…”