Structural-preserving speckle denoising for wafer surface phase images in digital holographic microscopy via hybrid NLM-enhanced U-Net CGAN

Digital Holographic Microscopy (DHM) offers irreplaceable advantages for high precision wafer surface defect detection, yet its 3D imaging is severely hampered by coherent speckle noise, particularly challenging on complex patterned surfaces where traditional methods prove inadequate. To overcome this, we propose an Enhanced Phase Denoising Network (EPDN), a novel CGAN that innovatively integrates NLM denoised features as an auxiliary input to an optimized U-Net generator, coupled with a tailored composite loss function. This CGAN-based framework effectively resolves the inherent conflict between noise removal and detail retention by meticulously suppressing speckle while preserving critical phase information without introducing blurring or over-smoothing. Our method demonstrates significant superiority over conventional and state-of-the-art approaches on simulated and real-world data, achieving a 14.67 % increase in GSNR, a 1.77 % improvement in Q-index, and a substantial 26.32 % reduction in phase error compared to the CGAN baseline. Crucially, spatial frequency analysis confirms that EPDN maintains the highest fidelity to the ground truth and exhibits a subtle yet measurable advantage over competitors in preserving high-frequency structural bandwidth. This robust approach provides reliable support for high-precision 3D surface profile reconstruction and showcases broad potential for industrial inspection.