Approach to weld segmentation and defect classification in radiographic images of pipe welds

Detecting indications (points suspicious for defects) in weld radiographs is an important research topic in the field of industrial non-destructive testing. Many computer-aided detection techniques have been designed for such applications as detecting indications occurrence, segmentation of the indication area, classification of the indications. However, these techniques are mainly focused on only one of the listed problems. Different defects may exhibit different visual properties in shapes, sizes, textures, contrasts, and positions, that often leading to ad-hoc solutions. The paper investigates to the fine tuning of the machine learning approach to high resolution radiograph processing produced by real-time radiography using digital detector arrays (DDA) method. The main contributions of this work are the preprocessing feature for human readability of the radiographic images and the proposed neural network-based solutions for all stages, from weld detection on the radiographic images and its segmentation to indication segmentation and classification. The designed approach was implemented as a web service with web site front-end. The demo version of the software is available with this instruction https://github.com/NastyaMittseva/DefectRecognitionSystem.