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

This paper presents a novel unsupervised hybrid approach combining Transformer and Mamba architectures with diffusion models for depth-dependent image denoising in Scanning Acoustic Microscopy (SAM). The method achieves state-of-the-art performance on SAM images for automated reliable inspection.

Abstract

Scanning Acoustic Microscopy (SAM) is a critical tool for non-destructive testing and semiconductor inspection. However, acoustic images often suffer from significant noise and artifacts that degrade inspection reliability. This work proposes an unsupervised hybrid Transformer–Mamba diffusion model to effectively denoise depth-dependent SAM images without requiring labeled training data. Our approach combines the global modeling capability of Transformers with the efficient state-space representations of Mamba, achieving superior denoising performance while maintaining computational efficiency.

Key Contributions

• Novel hybrid Transformer–Mamba architecture for acoustic image processing
• Unsupervised diffusion-based denoising framework
• Depth-dependent adaptation for multi-scale acoustic data
• Comprehensive evaluation on real SAM inspection datasets

Recommended citation: T. N. Nguyen *et al*., "Unsupervised Hybrid Transformer–Mamba Diffusion for Depth-Dependent Image Denoising in Scanning Acoustic Microscopy," in *IEEE Transactions on Industrial Informatics*, vol. 22, no. 4, pp. 3263-3274, April 2026, doi: 10.1109/TII.2025.3648838.