Physics-Guided Multi-Channel Restoration for Efficient and Accurate Atomic Force Microscopy Imaging

Scientific imaging instruments often acquire multiple physically complementary signals within a single measurement. In atomic force microscopy (AFM), high-throughput imaging requires higher scan speeds, however, image quality is fundamentally limited by multiple distortions. These artifacts are further amplified for sharp nano/microstructures, making accurate image restoration inherently challenging. Most deep learning-based approaches rely on a single primary channel and improve performance by increasing model complexity or training data. Here, we introduce a physics-guided multi-channel input approach for AFM that utilizes co-acquired auxiliary signals to improve image restoration efficiency and accuracy. We construct a four co-acquired channel input, each containing complementary physical information. By integrating these channels into a lightweight convolutional neural network, the proposed approach significantly improves restoration quality with minimal additional parameters. At a scan rate of 7.0 Hz, the four-channel DnCNN achieved higher reconstruction accuracy than a single-channel Restormer with approximately 47 times as many parameters, while providing approximately 24 times faster inference. Across the evaluated EFM and PFM conditions, four-channel inputs reduced the training-patch requirement by approximately two to three times at matched PSNR.