A multiscale attention network for mixed artifact suppression in AFM images

Accurate nanoscale characterization with Atomic Force Microscopy (AFM) is frequently hindered by complex mixed noise, particularly directional line artifacts and stochastic scars that stem from the instrument's electromechanical noise and feedback instability. While deep learning has shown promise, existing models often present an undesirable trade-off, either leaving residual artifacts or over-smoothing critical topographic details. In this paper, we propose a multiscale attention network named MS-HINet-CBAM, a multi-stage denoising framework specifically designed for AFM imaging. Our architecture integrates a Multiscale (MS) module to capture the long-range spatial correlations characteristic of mixed noises, and a Convolutional Block Attention Module (CBAM) to adaptively prioritize genuine topographic features over artifacts. Extensive evaluations on diverse simulated mixed noise scenarios demonstrate that our model consistently outperforms state-of-the-art methods in both visual fidelity and quantitative metrics. Analysis of the Root Mean Square (RMS) roughness reveals that while standard models significantly underestimate roughness, our model yields a result highly consistent with the ground truth, thus preserving the integrity of the surface. Furthermore, the practical utility of the proposed method is validated on real-world AFM images of organic molecules and bacterial cells, providing a robust tool for more reliable quantitative analysis in materials and life sciences.