Automated visual inspection is a cornerstone of modern smart manufacturing, yet deployment of deep learning-based defect detection systems on production lines remains constrained by the scarcity of labeled industrial data and the computational cost of edge inference without cloud connectivity. This paper presents a systematic benchmark of four fine-tuned YOLOv8 architectures (YOLOv8n/s/m/l) for real-time industrial surface defect detection on the NEU Surface Defect Database — 1,800 images across 6 steel surface defect categories. We fully train all four variants and evaluate the accuracy-efficiency trade-off under production-realistic constraints. YOLOv8l achieves the highest mAP@0.5 of 74.1% at 15.8ms inference on NVIDIA T4 GPU (43.6M parameters), while YOLOv8n achieves 73.6% at 2.1ms — a difference of only 0.5 percentage points at 7.5× the speed. Per-class analysis reveals substantial variance: patches (91.8%) and scratches (84.4%) are detected reliably across all variants, while crazing (50.6%) and rolled-in scale (55.6%) remain challenging due to their diffuse, texture-distributed morphology poorly suited to bounding box formulation. We further introduce an automotive manufacturing defect taxonomy covering body panel, weld, paint, and EV battery cell categories and outline a transfer learning pathway to automotive deployment. Our results establish a reproducible performance baseline for edge-deployed industrial defect detection and form the computer vision foundation for the ClosedMfgAI closed-loop manufacturing intelligence system.
Ezeji et al. (Sun,) studied this question.