Adaptive filtering on KPFM data: a deep neural network approach

Abstract

Kelvin probe force microscopy (KPFM) measurements often exhibit widely varying signal-to-noise ratios (SNR) because of non-ideal acquisition conditions, especially in time-sensitive experiments. Although conventional filters can suppress noise, inappropriate parameter settings may distort quantitative surface-potential information, and manual tuning reduces reproducibility. In this study, we propose a self-supervised deep learning framework for automatic optimization of filter parameters for KPFM data. The framework combines a multi-layer perceptron that predicts global filter parameters from 22 image-level features and structural complexity information with a convolutional neural network that generates pixel-wise local parameter adjustments. The model was trained on 13 218 real-world KPFM images without requiring clean reference data and adaptively adjusted filtering strength according to the noise level of each image. Under high-noise conditions, it achieved effective noise reduction while preserving structural features, yielding an edge preservation index of 0.984. Under low-noise conditions, it minimized unnecessary alteration of the original data, with a peak SNR of 39.4 dB and a structural similarity index of 0.940. The framework was implemented for five classical filters, namely Wiener, Gaussian, mean, bilateral, and total variation filters, and provides an objective and reproducible alternative to expert-dependent manual parameter tuning for quantitative KPFM analysis.