CrossMicroNet: a cross-scale small-sample image restoration framework for two-dimensional material microscopy imaging

High-quality microscopy is central to resolving the growth behavior, morphology, lattice organization and defect landscape of two-dimensional (2D) materials. Yet microscopy data acquired across scales are degraded by fundamentally different mechanisms: in situ optical microscopy (OM) is often compromised by motion blur, defocus, vibration-induced smearing and illumination inhomogeneity, whereas scanning transmission electron microscopy (STEM) is strongly affected by beam-induced amorphous carbon contamination. Here we introduce CrossMicroNet, a unified cross-scale image-clarification framework that couples a shared restoration front end with modality-adaptive refinement. The restoration module integrates contrast-limited adaptive histogram equalization, mild non-local means denoising, blind deconvolution, ringing suppression and wavelet-domain spatial-channel enhancement. For OM, this front end directly sharpens domain boundaries from small-sample data without requiring paired optical ground truth. For STEM, the restored output is further processed by a lightweight contamination-suppression branch that combines conservative structural guidance with smooth fusion to attenuate diffuse background haze while preserving lattice periodicity. Evaluated on 28 OM/SEM region pairs, where SEM serves only as an approximate structural reference, CrossMicroNet reduces the apparent edge-transition width to 0.22 μm and yields the most favorable overall trade-off among NIQE, LPIPS and PSNR-like structural-reference metrics. On the STEM benchmark, it achieves the best learning-based performance, with a PSNR of 20.50 dB, SSIM of 0.85, LPIPS of 0.08, VIF of 0.92 and FSIM of 0.93. Fourier-domain analysis further confirms suppression of low-frequency contamination while retaining lattice-frequency features. These results establish CrossMicroNet as a practical cross-scale clarification strategy for linking growth-scale OM with atomic-scale STEM in 2D-material research.