An Efficient Deep Neural Network for Surface Defect Detection in Industrial Edge Sensing

This article provides an efficient edge-end implementation solution for deep learning-based surface defect detection to improve the accuracy and efficiency when applied on edge devices with limited resource. An efficient you only look once (YOLO) network YOLOv5s-GhostNet is proposed, which highlights at the lightweight backbone/neck network, efficient feature extraction modules, and a fast learning scheme based on knowledge distillation. The parameter compression ratio is theoretically analyzed to show the decrease of computation complexity. The jointed loss is designed to enhance the generalization ability for new defects. An industrial testing platform with real-time edge-terminal-cloud detection system is developed with Raspberry Pi as edge. The experimental results show that the proposed method gets performances at complexity (floating-point operations per second (FLOPS) 8.2G, pt 7.9M), detection accuracy (precision 97.91$\%$, mean average precision (mAP) 96.66$\%$), efficiency [frames per second (FPS) 294 for single defect], and fast learning convergence (50 epochs). Compared to the existing methods, it reduces model size by 50$\%$ on overage, increases the detection efficiency by 4 times and maintains the higher accuracy.