Hybrid Metaheuristic-Optimized YOLO for Edge-Based Industrial Defect Detection in Smart Factory Systems
Year: 2026
Type: conference-paper
Abstract: Automated visual inspection of PCBs is a core requirement in smart manufacturing quality control, as even a hairline defect on a board can cause widespread failures at deployment scale. YOLO-based detectors satisfy the speed demands of modern production lines, yet their effectiveness is strongly tied to hyperparameter choices that are difficult to tune when accuracy and latency must be balanced simultaneously. This paper introduces HHO-YOLO, a framework that applies Harris Hawks Optimization (HHO) to YOLOv8s through a twoterm fitness criterion, allowing automated search across five training parameters while explicitly trading detection quality against inference cost. On a public PCB benchmark with six fault categories, HHO-YOLO achieves 98.95% [email protected] at 25 ms per image, surpassing Faster R-CNN (94.27%, 85 ms) and reaching accuracy comparable to TDD-Net (98.90%) at 2.8× lower latency. All reported values are means over three independent runs and are backed by per-class AP breakdowns, a confusion matrix, and an accuracy-latency scatter plot. The results show that systematic metaheuristic-driven configuration search is a lightweight, architecture-agnostic path to jointly better detection quality and faster response for edge-ready industrial deployment.
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202200147