Scholarlink

<p>Industrial digital radiography images are degraded by scattering that obscures defects, and conventional enhancement methods often sacrifice either noise suppression or detail preservation. This work presents a physics-inspired<b>,&#xa0;</b>annotation-free framework that explicitly separates the scattering component from the detected intensity through a radiation–matter interaction model. A multistage pipeline consisting of attenuation estimation, spatially adaptive scatter modeling, residual scatter removal, edge sharpening, and local contrast enhancement recovers the direct transmission signal. Evaluated on 60 industrial weld radiographs (ship plates, boilers, oil pipelines), the proposed method performs favorably against global histogram equalization, contrast-limited adaptive histogram equalization, a discrete wavelet transform approach, and a generic convolutional neural network baseline (serving as a lower-bound reference). It achieves the highest average contrast-to-noise ratio among the compared methods&#xa0;(up to 2.94), the lowest naturalness image quality evaluator scores (down to 4.67), and the lowest blind/referenceless image spatial quality evaluator scores (down to 22.37). Ablation studies verify incremental contributions of each stage, while sensitivity analyses on an independent test set and cross-condition validation confirm stable performance under parameter perturbations and across weld categories.&#xa0;The empirically calibrated coefficients are not derived from first principles, and cross-dataset generalization remains a necessary future investigation.&#xa0;The proposed framework provides a new detail-enhancement solution that significantly improves defect visibility and structural fidelity for industrial nondestructive evaluation using X-ray imaging. The C# source code developed for this study is freely available at <a href="https://doi.org/10.5281/zenodo.20711760">https://doi.org/10.5281/zenodo.20711760</a>.</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

Physics-inspired enhancement framework for industrial digital radiography based on radiation–matter interaction modeling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Fayu Chen,
  • Guancheng Lu,
  • Wei Wei

摘要

Industrial digital radiography images are degraded by scattering that obscures defects, and conventional enhancement methods often sacrifice either noise suppression or detail preservation. This work presents a physics-inspired, annotation-free framework that explicitly separates the scattering component from the detected intensity through a radiation–matter interaction model. A multistage pipeline consisting of attenuation estimation, spatially adaptive scatter modeling, residual scatter removal, edge sharpening, and local contrast enhancement recovers the direct transmission signal. Evaluated on 60 industrial weld radiographs (ship plates, boilers, oil pipelines), the proposed method performs favorably against global histogram equalization, contrast-limited adaptive histogram equalization, a discrete wavelet transform approach, and a generic convolutional neural network baseline (serving as a lower-bound reference). It achieves the highest average contrast-to-noise ratio among the compared methods (up to 2.94), the lowest naturalness image quality evaluator scores (down to 4.67), and the lowest blind/referenceless image spatial quality evaluator scores (down to 22.37). Ablation studies verify incremental contributions of each stage, while sensitivity analyses on an independent test set and cross-condition validation confirm stable performance under parameter perturbations and across weld categories. The empirically calibrated coefficients are not derived from first principles, and cross-dataset generalization remains a necessary future investigation. The proposed framework provides a new detail-enhancement solution that significantly improves defect visibility and structural fidelity for industrial nondestructive evaluation using X-ray imaging. The C# source code developed for this study is freely available at https://doi.org/10.5281/zenodo.20711760.