Edge-Deployed Deep Learning for Automated Quality Control in Industrial Assembly: A Case Study in Real-Time Defect Detection

Abstract

Recent advancements in computer vision (CV) and deep learning (DL) have significantly enhanced automated quality control in manufacturing. The use of deep Convolutional Neural Networks (CNNs) has revolutionized inspections by offering fast, accurate, and cost-effective solutions that minimize human intervention. This study focuses on a visual inspection task for quality control (QC) on an assembly line that is powered by a YOLOv8-based DL algorithm and integrates a robotic process control on the line. Through experimental evaluation, two YOLOv8 are trained and evaluated on a dataset containing annotated images collected from an active assembly line: an optimized YOLOv8.2s model and a verification YOLOv8.2m model. The optimized model achieves mAP50 of 0.972 and mAP50-95 of 0.647 across all classes, while the verification model shows marginal improvement with mAP50 of 0.979 and mAP50-95 of 0.673. However, this comes at the cost of noticeable increased computational demands, with training time increasing from 14 hours to 46 hours. The findings demonstrate that while larger models may offer slight performance improvements (2.6-3.8% in mAP50-95), the trade-off between computational resources and performance gains must be carefully considered in industrial applications.