LIGHTWEIGHT INDUSTRIAL DEFECT DETECTION ALGORITHM FOR EDGE COMPUTING: YOLOV8 OPTIMIZATION BASED ON KNOWLEDGE DISTILLATION

Abstract

Industrial defect detection faces significant challenges in deploying sophisticated deep learning models on resource-constrained edge devices. This paper presents a novel knowledge distillation framework for optimizing YOLOv8 to enable real-time defect detection in manufacturing environments. We propose a multi-level distillation approach incorporating response-based, feature-based, and attention transfer mechanisms specifically tailored for industrial defect characteristics. The framework systematically transfers knowledge from a full-capacity YOLOv8l teacher model (43.7M parameters) to a lightweight student model (3.8M parameters), achieving 91.3% parameter reduction while maintaining 98.7% detection accuracy. Our method addresses the unique challenges of edge deployment through hardware-aware optimizations, including operator fusion, quantization-aware training, and platform-specific kernel tuning. Comprehensive evaluation on NEU-DET, PCB defect, and fabric defect datasets demonstrates superior performance compared to existing lightweight architectures, with the distilled model achieving 84.6% mAP@0.5 while operating at 52.3 FPS on Jetson Nano. The framework outperforms alternative compression techniques, showing only 1.3% accuracy degradation compared to 12.4% for structured pruning and 7.8% for post-training quantization. Industrial deployment analysis reveals favorable cost-benefit ratios with 8-14 month return on investment. Memory consumption remains under 500MB across all tested edge platforms, enabling concurrent application execution. This research establishes new benchmarks for industrial edge AI deployment, providing both theoretical insights into knowledge transfer mechanisms and practical solutions for real-world quality control applications.