Deep Learning-Based Quality Inspection in Automated Production Lines

Abstract

Automated visual quality inspection is a critical requirement for modern high-throughput manufacturing, where manual inspection bottlenecks limitproduction speed and introduce subjective variability. Deep learning-basedcomputer vision systems have demonstrated transformative potential fordetecting surface defects, dimensional deviations, and assembly errors withsuperhuman accuracy and consistency. This paper presents the developmentand industrial validation of a real-time deep learning quality inspectionsystem deployed on a high-speed automotive stamping production line. Theproposed system integrates a multi-camera machine vision array with aYOLOv8-based object detection model fine-tuned for seven defect categories: scratches, dents, cracks, burrs, surface contamination, incomplete forming, and dimensional out-of-tolerance. The model was trained on a custom dataset of 28,400 annotated images captured under controlled industrial lighting conditions. Industrial validation on a production line operating at 42 strokes per minute demonstrated an overall defect detection accuracy of 97.6%, with precision of 96.8%, recall of 95.4%, and a false positive rate of 1.8%. Inference was executed on an NVIDIA Jetson AGX Xavier edge platform at 38 frames per second, enabling 100% inline inspection without cycle time impact. The system reduced escaped defect rate from 2.4% (manual inspection baseline) to 0.18%, representing a 92.5% improvement in outgoing quality. The findings confirm that deep learning vision systems can deliver production-grade quality inspection performance suitable for deployment in high-speed manufacturing environments [1], [2].  KEYWORDS: Deep Learning, Quality Inspection, Computer Vision, YOLOv8,Defect Detection, Automated Production, Surface Inspection, Edge Computing, Manufacturing Quality, Machine Vision