StripePoint-YOLO: Task-Adaptive Detection of Multi-Type Weld Seam Keypoints in Noisy Industrial Welding Scenes

Area of Science:

• Computer Vision
• Robotics
• Industrial Automation

Background:

• Weld seam keypoint detection is crucial for industrial automation.
• Complex interferences like arc light, spatter, and occlusion hinder stable detection.

Purpose of the Study:

• To develop a lightweight and robust model for weld seam keypoint detection.
• To improve accuracy and stability under challenging industrial conditions.

Main Methods:

• Proposes StripePoint-YOLO, a detection-based keypoint localization model built on YOLO11n.
• Introduces a P2 detection head, SPDConv, and removes the P5 detection output layer for enhanced feature representation.
• Utilizes WIoU v3 and NWDLoss for stable bounding box regression and an online physics-driven data augmentation (OPDDA) strategy.

Main Results:

• Achieves 80.46% mAP@50-95 and a Mean Center Error (MCE) of 2.33 px.
• Demonstrates a lightweight design with 1.86 M parameters and 19.42 GFLOPs computational cost.
• Reaches an inference speed of 159.80 FPS, showing stable keypoint localization across various weld seam types and noisy scenarios.

Conclusions:

• StripePoint-YOLO effectively addresses challenges in weld seam keypoint detection.
• The model offers a balance of accuracy, efficiency, and robustness for industrial applications.
• Validates the method's stability and effectiveness in complex, real-world welding environments.