Abstract

Several factors such as low image quality, limited human eye resolution, and biased interpretation make the manual radiography testing defect detection erroneous. The aim of this paper is to design a semi-automatic but accurate feature extraction and classification framework, using radiography images of welded joints represented in the GDXray image database. We have trained our support-vector machine classifier with crack, porosity and lack of penetration as the three more frequent classes of radiography defects. The images are segmented after binarization followed by a two-stage image enhancement technique. The body of the two-stage method is made up of an anisotropic diffusion Gaussian filtering, morphological edge detection, and low-pass Gaussian filtering. A three-class support-vector machine based on One-vs.-All implementation of the binary support-vector machine is created, trained, and tested. The method also involves manual adjustments which makes it semi-automatic, and compared to other studies, it is proved to work best with GDXray, a freely available comprehensive radiography image database. The effect of incorporating different optimization algorithms for solving the inherent optimization problem in support-vector machine theory, and utilizing various image processing techniques on defect detection is studied. We show that the combination of image processing and support-vector machine would result in a better performance than the previous studies using the same database.

中文翻译：

---

各向异性扩散图像增强与支持向量机相结合的基于补丁的焊缝缺陷分割与分类

摘要

诸如低图像质量，有限的人眼分辨率和偏倚的解释等多种因素使手动X射线摄影检测缺陷检测变得错误。本文的目的是使用GDXray图像数据库中表示的焊接接头的放射线图像设计一个半自动但准确的特征提取和分类框架。我们已经对我们的支持向量机分类器进行了训练，使其具有裂纹，孔隙率和穿透力不足等三种放射线缺陷类别。对图像进行二值化后再进行分段，然后采用两阶段的图像增强技术。两步法的主体由各向异性扩散高斯滤波，形态学边缘检测和低通高斯滤波组成。基于One-vs的三类支持向量机。-创建，训练和测试二进制支持向量机的所有实现。该方法还涉及手动调整，这使其成为半自动的，并且与其他研究相比，事实证明，该方法与GDXray（可免费获得的综合放射线图像数据库）一起使用时效果最佳。研究了采用不同的优化算法来解决支持向量机理论中固有的优化问题，并利用各种图像处理技术进行缺陷检测的效果。我们表明，与使用相同数据库的先前研究相比，图像处理和支持向量机的结合将产生更好的性能。事实证明，它与GDXray（可免费获得的综合放射线图像数据库）一起使用时效果最佳。研究了采用不同的优化算法来解决支持向量机理论中固有的优化问题，并利用各种图像处理技术进行缺陷检测的效果。我们表明，与使用相同数据库的先前研究相比，图像处理和支持向量机的结合将产生更好的性能。事实证明，它与GDXray（可免费获得的综合放射线图像数据库）一起使用时效果最佳。研究了采用不同的优化算法来解决支持向量机理论中固有的优化问题，并利用各种图像处理技术进行缺陷检测的效果。我们表明，与使用相同数据库的先前研究相比，图像处理和支持向量机的结合将产生更好的性能。