An automated appearance defect detection and removal system for varistors based on YOLO

To address the challenges of poor robustness, weak generalization capability in traditional machine vision methods for varistor defect detection, and the disconnection between detection algorithms and physical removal, this study proposes an automated detection and removal system for varistor appearance defects based on YOLOv8s. The system integrates an image acquisition module, a conveying mechanism, a cutting mechanism, and a control system, establishing a complete closed-loop process from visual perception to physical execution. The conveyor transports the varistor tape, while an industrial camera captures images in real-time. A lightweight YOLOv8s algorithm accurately identifies and locates defective varistors online, ultimately driving the cutting mechanism to remove them in real-time. Experimental results demonstrate the excellent detection performance of the system, achieving an overall mAP@0.5 of 0.870 on the test set and reaching the optimal detection precision (1.0) at a confidence threshold of 0.95. Furthermore, the F1-score remained consistently high across a wide confidence threshold range of 0.1 to 0.7, indicating strong robustness of the model. Ablation studies further revealed that a combined strategy of color enhancements (brightness, hue, exposure) and morphological enhancements (flipping, rotation, translation) effectively balanced the detection accuracy for defects of different sizes while mitigating overfitting risks. The full enhancement group (Group G) demonstrated the best overall performance balance across all five defect categories. This research provides a practical and effective solution for high-precision, automated online quality inspection of varistors.