A Generic Deep Learning Framework for Automated Vehicle Inspection

Ensuring consistent and precise vehicle inspection is important in automotive manufacturing. Manual inspection often suffers from variability, subjectivity, and limited scalability. Recent advances in deep learning now allow for automated visual inspection. However, most current systems focus on specific defect categories and act as "black-box" models. This limits their adaptability and reliability in real-world situations. This paper presents GVI-Net, a generic and explainable deep learning framework for inspecting multiple vehicle components. The framework uses a shared ResNet-50 backbone with lightweight task-specific classification heads. This setup allows for scalability across various inspection tasks, such as detecting tire wear, identifying windshield cracks, and assessing bumper alignment. Grad-CAM-based visualizations give interpretable feedback to human inspectors. In addition, a human-in-the-loop mechanism enables ongoing model refinement based on actual inspection results. Experimental evaluations on a hybrid dataset of real and artificially enhanced images show strong performance across all inspected components. There were also measurable improvements after interactive retraining. The findings suggest that GVI-Net is suitable for industrial use, enhancing inspection consistency, transparency, and adaptability.