Scholarlink

Background <p>Deep learning-based digital image correlation (DL-DIC) has emerged as a new paradigm for image deformation measurement, offering a promising and powerful solution to circumvent the limitations of traditional DIC. However, all data-driven deep neural networks used in existing DL-DIC lack a metric to quantify the matching quality at each pixel point, and therefore cannot guarantee reliable or provably correct results. How to assess the image matching quality of DL-DIC and the metrological performance of various DL-DIC networks is therefore a critical and fundamental problem towards its practical use.</p> Objective <p>In this work, we propose a simple and effective metric called squared intensity residual (SIR) to assess the pixel matching quality of DL-DIC.</p> Methods <p>SIR is defined as the squared intensity difference (or squared&#xa0;gray level residual) of matched pixel points in the reference and deformed images. Also, to mitigate the adverse influence of inevitable image noise and ambient illumination variations on SIR calculation, the reference and target images are first denoised using Gaussian low-pass filtering, followed by the application of an illumination coefficient correction scheme.</p> Results <p>Experimental results demonstrate that SIR reliably assesses matching quality and effectively quantifies spatial error distribution of DL-DIC outputs.</p> Conclusions <p>The proposed SIR can serve as an objective, fair and practical metric for evaluating the metrological performance of various DL-DIC methods, thereby laying&#xa0;a solid foundation for practical DL-DIC applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Squared Intensity Residual: A Simple and Effective Matching Quality Assessment Metric for Deep Learning-Based Digital Image Correlation

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Y. Liu,
  • K. Qian,
  • B. Pan

摘要

Background

Deep learning-based digital image correlation (DL-DIC) has emerged as a new paradigm for image deformation measurement, offering a promising and powerful solution to circumvent the limitations of traditional DIC. However, all data-driven deep neural networks used in existing DL-DIC lack a metric to quantify the matching quality at each pixel point, and therefore cannot guarantee reliable or provably correct results. How to assess the image matching quality of DL-DIC and the metrological performance of various DL-DIC networks is therefore a critical and fundamental problem towards its practical use.

Objective

In this work, we propose a simple and effective metric called squared intensity residual (SIR) to assess the pixel matching quality of DL-DIC.

Methods

SIR is defined as the squared intensity difference (or squared gray level residual) of matched pixel points in the reference and deformed images. Also, to mitigate the adverse influence of inevitable image noise and ambient illumination variations on SIR calculation, the reference and target images are first denoised using Gaussian low-pass filtering, followed by the application of an illumination coefficient correction scheme.

Results

Experimental results demonstrate that SIR reliably assesses matching quality and effectively quantifies spatial error distribution of DL-DIC outputs.

Conclusions

The proposed SIR can serve as an objective, fair and practical metric for evaluating the metrological performance of various DL-DIC methods, thereby laying a solid foundation for practical DL-DIC applications.