A Flexible Multitask Framework With Decoupled Training Paradigm for Radiographic Weld Images

Weld sign detection, weld edge segmentation, and weld defect detection are three major tasks in Artificial-Intelligence empowered radiographic weld images. Most of existing methods focus on improving detection performance on single task, the others focus on two tasks, and currently still lacking a more comprehensive detection method. We argue that simultaneously accomplishing three tasks can not only efficiently utilize the multi-task measurement data for downstream industrial applications, but also enhance the resource utilisation of industrial deployment. To this end, we propose a flexible multi-task framework for effectively addressing three tasks. First, a multi-task framework is elegantly designed, where the shared features is used as a flexible interface for connecting shared feature encoder with multiple task-specific prediction heads. Particularly, a two-stage decoupled training method is proposed to improve multi-task method’s accuracy by separately pre-training the shared feature encoder and fine-tuning the task-specific prediction heads. We instantiate our framework on an open-source dataset GDXray and a real-world dataset. The experimental results demonstrate that the proposed method achieves a higher balance of accuracy and efficiency on three tasks, including weld defect detection (0.959 mAP, mean average precision), weld edge segmentation (0.941 mAP), and weld sign detection (0.972 mAP).