Smart-RT : End-to-end automatic radiographic images analysis through a multi-stage object detection algorithm

A meticulous inspection of manufactured parts is required in any production chain as well during the operations cycle to ensure the quality of the produced and operating equipment. Among those inspection steps, non-destructive evaluation (NDE) of welds is a key point of industrial testing as some manufacturing segments (nuclear, aerospace, oil and gas industry) need their parts to resist severe conditions. Weld quality can be assessed through numerous means such as eddy current, ultrasound or X-ray testing. This paper focuses on the latter and introduces a multi-stage neural network-based algorithm for a fully automated radiographic image quality assessment. Not only does the quality of the weld needs to be checked to ensure that no defects are introduced during the welding process, but the quality of the digital X-ray image itself has to undergo specific criteria (marking, optical density range, visible image quality indicators...) to be considered as exploitable. To this end, a diverse dataset of X-ray images annotated by radiographic testing experts for training and validation purposes. Various advanced object detection algorithms were applied and compared across key image analysis tasks: optical density assessment, marking identification, image quality indicators (IQI) recognition and defects detection. All those steps are integrated into a comprehensive framework, enhancing the efficiency and accuracy of X-ray image analysis.