Blur-guided conditional diffusion for robust correlation in fast-scanned digital images of in-situ SEM tensile testing

Abstract

This study proposes a blur-guided conditional diffusion framework for restoring noisy in-situ scanning electron microscopy (SEM) images acquired with a reduced number of scan accumulations for accurate measurement of material properties, focusing particularly on structural and edge preservation under fast-acquisition conditions. During in-situ SEM tensile testing, reducing scan accumulations significantly increases image noise, degrading the accuracy of digital image correlation (DIC) based displacement measurements. Therefore, for accurate metrological analysis, image restoration must prioritize DIC applicability by preserving the structural features required for measurement, rather than conventional denoising. To address this, the proposed method formulates a diffusion-based restoration process within a conditional framework guided by a blur condition. This condition preserves the low-frequency shape and positional alignment from the observed image, while the reverse diffusion process reconstructs high-frequency boundaries and textures essential for reliable DIC measurement. Furthermore, an ensemble inference strategy is incorporated to suppress stochastic sampling variations, thereby ensuring high restoration stability. Experimental results with real SEM data demonstrate that peak signal-to-noise ratio is improved from 11.67 dB for noisy inputs to 18.10 dB for the proposed method. In full-field DIC validation across noise levels ( σ = 0.10–0.50) and imposed shifts (0.5–3.0 px) against classical filters, a learning-based U-Net baseline, and the unprocessed input, the proposed method is the only approach that preserves sub-pixel field accuracy at moderate noise (0.79 px field mean absolute error at σ = 0.30 for a 0.5 px shift, versus 1.27 px for the unprocessed input and 2.23 px for U-Net), degrades most slowly under severe noise (up to 44% lower error than the unprocessed input at σ⩾ 0.4), and reduces the variance of the displacement estimate by a factor of approximately 3–8 relative to all competing restoration methods, reflecting a bias–variance trade-off characterized in this work. Consequently, the proposed framework enhances the reliability of deformation tracking in high-speed, high-noise in-situ SEM environments, establishing a practical strategy for metrological SEM image restoration under severe acquisition noise.