Fig 1.
Defects of electrode cap tip surface in multi-scale and high intra-class variability.
Black burn marks that are: (a) tiny dots, (b) elongated stripe, (c) scattered, (d) partial patch, (e) full patch, (f) curved stripe at edge.
Fig 2.
Representative samples of burn marks in the ECTSD dataset show differences in sizes and shapes.
Fig 3.
Network architecture of YOLOv13.
Fig 4.
BSAM block.
Fig 5.
BiLevel routing attention module.
Fig 6.
Integration of the BSAM module into the: (a) backbone of YOLOv13, (b) neck of YOLOv13, (c) head of YOLOv13.
Fig 7.
PConv block.
Fig 8.
The overall network architecture of YOLO-BP.
Table 1.
Performance comparison of BSAM module inserted at three positions: Backbone, Neck, and Head.
Table 2.
Results of ablation experiments.
Fig 9.
Comparison of detection results over epochs on the ECTSD Dataset.
Fig 10.
Variability comparison across independent training runs for performance metrics of models at 50-epoch intervals.
Data points and error bars indicate mean and standard deviations respectively.
Fig 11.
Detection result of electrode cap tip surface defect dataset.
Table 3.
Results of comparative analysis between YOLO-BP and other models.