Coarse-to-fine vision-based welding spot anomaly detection in production lines of body-in-white

Abstract

Computer vision-assisted methods for weld quality inspection enable rapid and automated surface defect detection through image data. However, the application of Computer Vision-Assisted Inspection (CVAI) in real-world production lines faces substantial, long-term challenges due to complex environments, imbalanced data samples, real-time processing demands, and safety requirements. Our paper proposes a novel two-stage Coarse-to-Fine Anomaly Detection (CTFAD) framework, which integrates the YOLOv8 network architecture for initial detection with an ensemble of neural networks for fine-grained classification. Additionally, we introduce a voting-based algorithm for improved decision-making accuracy. Experimental results on real-world datasets demonstrate that, compared to standard end-to-end methods, CTFAD enhances detection accuracy and operational efficiency. Our contributions include (1) proposing the CTFAD pipeline for weld anomaly detection, (2) establishing voting-based classification module to increase system robustness and generalization, and (3) developing an integrated weld detection system encompassing data acquisition, processing, analysis, and anomaly alerting. Our code is available at https://github.com/wj-liu0730/ctfad-jms.