SPAD-Net: Structure-Preserving Photoelectron- Adaptive Denoising Network for Low-Exposure ARPES Image

Abstract

Low-exposure angle-resolved photoemission spectroscopy (ARPES) images are strongly affected by Poisson-Gaussian noise, which degrades band boundaries and peak shapes and reduces the reliability of physical interpretation. However, most existing methods primarily perform denoising by minimizing pixel-wise errors. As a result, methods that substantially improve denoising performance while facilitating reliable structural interpretation remain limited. In this paper, we propose a structure-preserving photoelectron-adaptive denoising network (SPAD-Net) for low-exposure ARPES images by integrating a decoder-attention-enhanced U-Net, noise-regime-adaptive sampling, and a structure-preserving loss. The SPAD-Net updates sampling probabilities online according to regime-wise relative reconstruction errors to reduce training bias across noise regimes. In addition, a combined loss design with absolute-error (L1), gradient, and energy distribution curve (EDC)-consistency terms jointly promotes pixel-error reduction and band-structure preservation. Under a controlled Poisson–Gaussian noise model, the SPAD-Net achieved an average peak signal-to-noise ratio (PSNR) of <inline-formula> <tex-math notation="LaTeX">$32.92\pm 0.23$ </tex-math></inline-formula> dB and an average multiscale structural similarity (MS-SSIM) of <inline-formula> <tex-math notation="LaTeX">$0.9758\pm 0.0016$ </tex-math></inline-formula> across nine total photoelectron counts from <inline-formula> <tex-math notation="LaTeX">$N= 1.0\times 10 ^{5}$ </tex-math></inline-formula> to <inline-formula> <tex-math notation="LaTeX">$1.0\times 10 ^{7}$ </tex-math></inline-formula>. These results exceeded RUnet by 6.99 dB in PSNR and 0.0241 in MS-SSIM and ARPESNet by 9.40 dB in PSNR and 0.0381 in MS-SSIM. At <inline-formula> <tex-math notation="LaTeX">$N= 5.0\times 10 ^{6}$ </tex-math></inline-formula>, the SPAD-Net achieved mean absolute error (MAE) reduction rates of <inline-formula> <tex-math notation="LaTeX">$71.37\pm 0.25$ </tex-math></inline-formula>% for EDC and <inline-formula> <tex-math notation="LaTeX">$68.88\pm 1.88$ </tex-math></inline-formula>% for the momentum distribution curve (MDC). On public experimental WSe2 and Bi2212 data, the SPAD-Net achieved the highest PSNR among the evaluated ARPES-specific methods without additional training or fine-tuning.