Interactive defect segmentation in X-Ray images based on deep learning. (15th July 2022)

• Record Type: Journal Article
• Authors: Du, Wangzhe; Shen, Hongyao; Zhang, Ge; Yao, Xinhua; Fu, Jianzhong
• Is Part Of: Expert systems with applications. Volume 198(2022)
• Publication Date: 2022-07-15
• Subjects: Interactive defect segmentation -- Casting aluminum parts -- X-ray image -- Deep learning -- Nondestructive testing -- Computer vision; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.116692
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Intelligent defect detection systems based on deep learning have substantial potential to be applied in segmenting defects of aluminum casting parts' X-ray images. The success of data-based methods relies on well-established datasets that require time and labor to perform the annotation processes necessary to build them. In efforts to improve the efficiency of image annotation, many deep learning-oriented interactive segmentation types of research have been conducted with a single natural object that has high semantic information. However, these approaches are not suitable for defects with low semantic information in X-ray images because the defects have characteristics such as blurred boundaries, multiple scales, random distribution in one region, and low contrast. To fill this gap in the field of deep learning-based interactive segmentation for X-ray images, we conduct this study. First, static and dynamic point sampling strategies are proposed to generate foreground and background points to guide model training. Second, we propose an iterative training procedure to help the model perceive both global and point-specified information stage by stage. And third, an interactive X-ray network (IXNet) with a click attention module is proposed to achieve the interactive segmentation for the multiple casting defects with low semantic information in one region simultaneously. The experimental results prove the effectiveness of the proposed methods: the iterative training process improved by 13% and 5% in mNoC (75%) and mNoC (80%), respectively compared with the pure training process, and IXNet with the click attention module can achieve 2.38 and 3.96 in mNoC (75%) and mNoC (80%), respectively.