DARA-Net: a dual-path residual attention network for denoising AFM images under complex noise conditions

Abstract

Atomic force microscopy (AFM) images often suffer from random fluctuations, periodic interference, and scanning artifacts. We propose DARA-Net, a denoising framework combining a dual U-shaped GAN architecture with residual attention blocks (RABs) in the decoder stage. Two parallel predictors—one for structural content and one for noise—enable effective feature disentanglement, while the GAN framework enhances detail realism. The fused output benefits from the RAB's ability to suppress complex noise and preserve nanoscale topographical features. A dataset of 1000 in-house AFM images with synthetic degradation was used for training and evaluation. The results show that DARA-Net outperforms classical and state-of-the-art methods in PSNR, SSIM, and RMSE and achieves lower errors in four physical surface metrics (perimeter, height, roughness, and volume), demonstrating superior generalization and structural preservation for nanoscale imaging.