Smart Edge-AI Solution for Automated Real-Time Surface Defect Identification

Abstract

This paper presents a Smart Edge-AI framework for automated real-time surface defect identification using deep learning and computer vision. The system employs a lightweight YOLOv8n object detection model deployed on a distributed architecture consisting of a Raspberry Pi 4 for video acquisition and wireless streaming, and a host computer for deep learning inference. The framework incorporates dataset refinement techniques, including data augmentation, negative sample integration, and annotation verification, to enhance model robustness. Model optimization through ONNX conversion, input resolution tuning, and confidence threshold calibration further improves inference speed. The system achieves a detection precision of approximately 0.96, an F1 score of 0.93, and operates at 15–20 frames per second in real-time. The proposed solution is portable, cost-effective, cloud-independent, and suitable for deployment in infrastructure monitoring, industrial inspection, and public safety applications. Experimental results validate that the combination of edge computing and deep learning provides a scalable and efficient alternative to conventional inspection methods.