A Robust Deep Learning-Assisted Digital Image Correlation for Deformation Measurement at 1600 °C in Air

Digital image correlation (DIC) is an image-based deformation measurement method. However, problems such as heat haze, speckle oxidation and debonding, and image overexposure in ultra-high-temperature environments lead to image degradation and compromise the reliability of deformation measurement.

This study proposes a robust and high-precision DIC algorithm designed to measure deformation stably from low-quality speckle images by leveraging machine learning. An ultra-high-temperature in-situ X-ray imaging device addresses challenges like speckle instability and heat haze interference. The proposed algorithm and experimental device are combined to measure the deformation field at 1600 °C in air.

A novel image matching network-assisted digital image correlation (IMN-DIC) is proposed. This approach uses a deep learning-based image matching network to extract and match features for initial displacement estimation. Subsequently, an iterative algorithm based on the inverse compositional Gauss–Newton (IC-GN) method is applied to achieve sub-pixel accuracy in high-temperature deformation field measurements. Numerical experiments and real experiments of C/SiC composite samples under tension at 1600 °C in the air with optical and X-ray imaging were carried out to verify the effectiveness of the IMN-DIC.

For high-quality optical speckle images, IMN-DIC achieved comparable measurement accuracy but with greater computational efficiency than previous feature-based DIC methods. In X-ray images captured at 1600 °C in air, the traditional DIC method successfully processed only 50.17% of points of interest (POIs), whereas IMN-DIC achieved 98.96%, demonstrating superior robustness.

The IMN-DIC method exhibits high robustness, reliably capturing deformation data from low-quality speckle images with weak textures and high noise levels. This approach holds significant promise for applications in extreme environments where artificial speckle generation is challenging and image quality is compromised.