Machine Vision-Based Real-Time Surface Defect Detection using Deep Convolutional Neural Networks for High-Speed Automotive Production Lines: Architecture, Training, and Industrial Validation

Abstract

Automated visual inspection using deep learning- based computer vision systems has emerged as the predominant approach for achieving zero- defect manufacturing targets in high- speed production environments where manual inspection throughput and consistency constraints are prohibitive. This paper presents DefectNet-7, a custom seven- layer convolutional neural network architecture designed and optimised for real- time surface defect detection on automotive stamped metal components at line speeds up to 120 parts per minute. The network was trained on a dataset of 48,000 labelled images across six defect categories: scratches, dents, porosity, burrs, cracks, and conforming surfaces, captured under controlled multi- spectral illumination conditions. DefectNet-7 achieved 97.4% overall accuracy on the held- out test set, with per- class F1 scores ranging from 0.938 to 0.976, and mean inference time of 8.3 ms per image on NVIDIA Jetson Xavier NX embedded hardware. Industrial deployment across two production lines over six months resulted in a 91.3% reduction in escaping defect rate, USD 420,000 annual quality cost savings, and a 2.4- month payback period.