C3TR-DCN: an enhanced YOLOV8-S for automated defect detection in phased array ultrasonic testing of austenitic welds

Automated defect detection in PAUT of austenitic stainless-steel welds is hampered by significant anisotropic noise and irregularly shaped flaws, which are challenging for conventional methods to identify. Thus, an improved object detecting model C3TR-DCN YOLOv8-s is proposed in this paper: improving yolo-v8-s architecture so that for each suspicious thermal fatigue crack it gives not only the position information but also extent information and certainty level information. The whole image context awareness sub-model (C3TR) is provided by C3 Transformer; when encountering an anomaly in flaw shape, DCN V2 makes adaptive processing; WIoU Loss ensures robust training under background variation. On ML NDT dataset, this research obtained excellent test results mAP@0.5=97.1%, F1-Score=95.5%. Testing results show a 3.0 percentage point increase in mAP@0.5 compared to the baseline YOLOv8-s. This detector can quickly get the accurate box and fast inference speed too. A real-time reliable solution is offered for industrial applications using autonomous inspection techniques.