作者
Huiming Chen,Yaorong Huang,Haoran He,Shuai Xue,Xinya Gao,Xulin Zhang,Yahong Wang
摘要
Low-contrast radiographic weld defects are difficult to detect automatically because small and weakly bounded indications are easily confused with weld texture, image-quality markers and background intensity variation. This study presents ELD-YOLO, an explainable and edge-deployable YOLO framework for detecting cracks, lack of fusion, lack of penetration, porosity and slag inclusion in plate-weld radiographs. The framework first localizes the weld region and then performs overlapping sliding-window defect detection on local weld patches. To improve sensitivity to faint and small defects, a Dual Feature Cross Attention with Intensity Enhancement (DFCA-IE) module is embedded in YOLO11n to couple cross-level feature interaction with defect-related intensity enhancement, while Grad-CAM is used to inspect whether model responses align with defect regions. A dataset was constructed from 68 public GDXray images and 374 industrial radiographic weld images, producing 3748 augmented defect samples after annotation, weld-region extraction and augmentation. On the validation set, ELD-YOLO achieved a precision of 0.947, recall of 0.856, F1-score of 0.900, mAP@50 of 0.918 and mAP@50:95 of 0.627, with 4.02M parameters, 15.7 GFLOPs and 156.8 FPS on a workstation GPU. Compared with the best-performing detection-head configuration (P5), precision, recall and F1-score increased by 11.5%, 19.2% and 16.9%, respectively. Comparison of global and ROI-constrained Grad-CAM maps showed improved spatial concentration around correctly detected defects, while representative false-positive and missed-detection cases demonstrated the visualization’s utility for failure analysis.In an 8.92-h continuous Raspberry Pi 5 deployment test, the complete processing and recording workflow was executed on 5360 external radiographic images, with an average processing time of 5.99 s per image. Mean core detection and end-to-end times were 3.67 and 5.99 s, respectively, with 956.97 MB mean peak process memory and a mean observed temperature of 56.9 °C. These findings demonstrate stable long-duration standalone execution.