Accurate segmentation of weld defects with horizontal shapes. (March 2022)

• Record Type: Journal Article
• Authors: Radi, Doaa; Abo-Elsoud, Mohy Eldin A.; Khalifa, Fahmi
• Is Part Of: NDT & E international. Volume 126(2022)
• Publication Date: 2022-03
• Subjects: Weld defect -- Segmentation -- Background subtraction -- Computer-aided technique; Nondestructive testing -- Periodicals; Contrôle non destructif -- Périodiques; Electronic journals; 620.1127
• DOI: 10.1016/j.ndteint.2021.102599
• Languages: English
• ISSNs: 0963-8695

Abstract

Abstract: Many researchers used computer-aided detection techniques for weld flaws segmentation problems. However, existing approaches fail to segment different weld defects types. We elaborate on our recently published work [1] to provide a comprehensive framework capable of detecting weld defects. This paper proposes a novel image processing pipeline that explores two new filters to segment the horizontal flaws and separates the background and vertical defects. Our technique performs two separate convolutional processes between the preprocessed image and both filters. We test our method on a publicly available GDXray database, which consists of 68 weld X-ray images for nondestructive testing [2]. Experiments showed that the proposed technique achieved a high segmentation accuracy, based on multiple evaluation metrics such as Dice and Jaccard metrics. The efficiency of our method is high. The proposed pipeline demonstrated its efficiency and applicability (Dice and Jaccard metrics are >99%), in practical weld defect segmentation processes, despite its simple structure. Thus, it is a promising technique for the detection of weld defects. In total, our technique can segment more different flaws types that exist in weld radiographs, compared to traditional approaches for the detection of weld defects.