Unsupervised Hybrid Transformer–Mamba Diffusion for Depth-Dependent Image Denoising in Scanning Acoustic Microscopy

Scanning acoustic microscopy (SAM) plays a critical role in nondestructive inspection of semiconductor and industrial materials; however, deep-layer SAM images often suffer from severe attenuation-induced noise and the absence of clean reference data, significantly limiting reliable defect analysis. To address these challenges, this article proposes a hierarchical diffusion-based denoising framework that transfers structural and statistical knowledge from high-quality top-layer SAM images to restore degraded deep-layer images. A physics-guided synthetic data generation pipeline, combining Gaussian degradation modeling and unpaired domain transfer, is introduced to enable effective training. A hybrid Transformer–Mamba diffusion architecture is also constructed to jointly capture local spatial structures and long-range depth-dependent dependences, while a depthwise latent extraction strategy enhances the separation of noise and structural features. Extensive experiments on benchmark datasets and industrial SAM images demonstrate that the proposed approach achieves better performance on both reference-based metrics (peak signal-to-noise ratio, structural similarity index measure, learned perceptual image patch similarity, and Fréchet inception distance) and no-reference evaluation, while significantly improving downstream defect segmentation accuracy using the segment anything model. Although the proposed model has a higher parameter count than competing methods, the results indicate its strong potential for practical SAM-based industrial inspection.