ABSTRACT Maintaining high quality in automotive manufacturing is essential, as even small defects can lead to safety issues, costly recalls, and increased operational costs. Manual inspection is often unreliable in fast‐paced production, limited by human error and poor scalability. Advanced imaging and deep learning‐based Anomaly Detection and Localization (ADL) offer effective alternatives, but their use in industry is challenged by factors like complex geometries, inconsistent lighting, and environmental noise. This work presents an ADL framework for inspecting sealant application in car underbodies that combines a video acquisition system with four state‐of‐the‐art deep learning models. To overcome the lack of annotated data, a synthetic defect generation module is introduced, creating realistic anomalies that improve model evaluation while reducing annotation effort. The framework was tested on both synthetic and real‐world data, achieving high localization performance (AUROC up to 99.7%, F1‐score of 43.4%) with inference times ranging from 0.08 to 3.33 s depending on model complexity. These results highlight the trade‐offs between speed and accuracy, and confirm the potential of ADL models for real‐time quality control in industrial automotive settings.
Manigrasso et al. (Sun,) studied this question.