Subset-based Smart-DIC vs. deep learning-based DIC: Which is better?

Zhao, Jianhui

doi:10.1016/j.optlaseng.2026.109896

Subset-based Smart-DIC vs. deep learning-based DIC: Which is better?

2026-05-01

1

PRE

AI

Bing Pan (Bing Pan) *

DOI：10.1016/j.optlaseng.2026.109896

Original

Original request for help

Save

Abstract

En 中文

Digital Image Correlation (DIC) techniques rely on an image matching/registration/tracking algorithm to track corresponding physical points across different speckle images. Among various implementations, subset-based DIC algorithms have been extensively studied, continuously refined, and widely adopted in most commercial DIC software. Smart-DIC, representing the state-of-the-art in this category, achieves user-independent and accurate measurement via self-adaptive parameter selection. Recently, the rise of deep learning has opened a new frontier in the DIC community. Deep learning-based DIC (DL-DIC) methods enable pixel-wise and end-to-end displacement measurement. In some studies regarding DL-DIC, comparisons with subset-based DIC were limited to deformation accuracy and/or spatial resolution. Moreover, the subset-based methods in these studies require manual setting of key parameters, including subset size and shape function order, which introduces potential bias and limits the objectivity of the evaluation. Additionally, these DL-DIC networks are limited by accuracy, displacement range, and generalization capability. To address these limitations, G-RAFT was recently proposed to provide accurate DIC measurements with a large displacement range, representing the current stateof-the-art DL-DIC. In this work, we present the first systematic comparison of Smart-DIC and G-RAFT under controlled simulations and real-world experiments, benchmarking their performance in terms of noise robustness, speckle generalization, deformation accuracy, and computational efficiency. Experimental results demonstrate that Smart-DIC achieves higher accuracy when deformations are well approximated by shape functions and demonstrates stronger generalization to low-quality speckles, while G-RAFT excels in complex deformation scenarios and delivers significantly higher computational efficiency. These findings offer practical guidance for image matching algorithm selection in diverse applications and underscore the potential of hybrid approaches that combine the robustness of Smart-DIC with the modeling flexibility of G-RAFT.

Keywords:

Digital image correlation

Subpixel displacement measurement

Self-adaptive optimization

Deep learning

Subset-based Smart-DIC vs. deep learning-based DIC: Which is better?

Abstract

Journal

Organization

Cited Papers