3D digital image correlation (3D-DIC) is a pivotal non-contact full-field measurement method in experimental mechanics. Traditional 3D-DIC systems rely on multi-camera setups, which face challenges such as high hardware costs, complex synchronization, and limited stereo matching accuracy. Single-camera stereo vision mitigates synchronization issues but introduces inherent resolution loss, compromising measurement accuracy. To address this, we propose LaESR-Diff, a super-resolution diffusion model integrating an optimized enhanced super-resolution generative adversarial network (ESRGAN) with Laplacian noise scheduling. The LaESR-Diff model employs an optimized ESRGAN to generate conditional inputs, enhancing residual-in-residual dense blocks (RRDBs) and incorporating a zero-mean normalized cross-correlation (ZNCC) loss term for speckle image fidelity. A Laplacian noise schedule replaces traditional linear/cosine schedules to preserve high-frequency textures. A dual evaluation system combines general image metrics and 3D-DIC-specific accuracy metrics. Experiments on the “Stereo-DIC Challenge 1.0” dataset show LaESR-Diff achieves PSNR = 26.13 and SSIM = 0.7423 at × 8 scale, reducing surface height and displacement errors by 58.6% and 67.2%, respectively, compared to bicubic interpolation. Laboratory tensile tests confirm LaESR-Diff reduces full-field displacement error in a quadrangular-prism-based virtual stereo system (QVSS) from 4.25% to 1.71%, nearing the basic stereo system (BSS) accuracy. The results demonstrate that the proposed method effectively compensates for the accuracy degradation caused by resolution loss in single-camera systems. This study provides an effective solution for single-camera 3D-DIC accuracy enhancement, with potential applications in other image-based non-contact measurement fields.
Zhou et al. (Thu,) studied this question.