Deep-learning-assisted scattering structured-illumination confocal microscopy for industrial super-resolution imaging

Laser scanning confocal microscopy (LSCM) plays an essential role in industrial inspection, yet its performance is fundamentally limited by the "edge shadow effect" in reflection imaging and the Abbe diffraction limit. To address these challenges, we propose a deep learning-driven scattering laser scanning structural illumination microscope (DL-sLSSIM) for super-resolution inspection of semiconductor nanostructures. A polarization-compensated scattering confocal configuration is first established to suppress specular reflection background and enhance edge contrast in steep topographies. We further analyze the imaging mechanism of point-scanning structured illumination and show that morphology-dependent reflection-scattering modulation induces severe stripe distortion, which prevents conventional SIM reconstruction in industrial samples. To overcome this issue, a residual channel attention network (RCAN) is introduced to perform nonlinear demodulation of the distorted structured light patterns and accurately extract the high-frequency image information from the distorted fringes. Experiments on semiconductor wafers demonstrate that DL-sLSSIM achieves a lateral resolution improvement of approximately 1.8×, reaching 480 nm under low-magnification inspection conditions, while significantly improving the signal-to-noise ratio. The proposed method provides a promising route for high-fidelity, non-destructive super-resolution metrology in advanced semiconductor manufacturing.