DIPSAM: Diffusion inpainting for restoring water-loss degradation in scanning acoustic microscopy

Scanning Acoustic Microscopy (SAM) is a widely used ultrasonic inspection technology in non-destructive testing (NDT), offering high-resolution imaging for defect detection in industrial and biomedical applications. However, signal distortions caused by water loss in the coupling medium introduce missing regions in acquired images. These image degradations compromise defect visibility and lead to reduced accuracy in automated segmentation, thereby increasing the risk of undetected failures. To address this, we propose Diffusion-based Image Inpainting for Scanning Acoustic Microscopy (DIPSAM) to reconstruct corrupted regions in SAM images affected by water-loss degradation. Building on existing diffusion achievements, our DIPSAM model employs a two-stage latent diffusion process to optimize computational efficiency while preserving high reconstruction fidelity. A unique line mask generation strategy, tailored to the characteristics of SAM images, is introduced to create realistic training datasets. Experimental results demonstrate that the proposed method significantly improves image reconstruction quality and increases defect segmentation performance. The restored images achieve a higher Peak Signal-to-Noise Ratio (32.96 decibels), a lower Learned Perceptual Image Patch Similarity (0.016), and an improved Dice Similarity Coefficient (up to 0.943) compared to state-of-the-art baselines. This work not only advances diffusion-based SAM image restoration but also provides valuable insights into designing efficient computational frameworks for degraded ultrasonic and related imaging systems.