Real-time defect detection and classification in energy cable production lines using an improved YOLOv8 algorithm

In this study, a deep learning-based detection and classification network architecture was enhanced for the real-time detection and classification of defects occurring on cable surfaces during the production of energy cables. The proposed approach was improved as a hybrid architecture that simultaneously incorporates Focus, Ghost Convolution and C3TR modules. To create a real-time dataset, a prototype cable production line was designed and developed. Using this prototype line, a real-time dataset consisting of three different defect types Fisheye, Dent, and Slit was created. Following this, training phases were conducted using various algorithms, including YOLOv5-v8-v9-v11-v12, improved YOLOv8, and Faster R-CNN. The performance of each algorithm was evaluated using metrics such as precision, recall, mean average precision (mAP@50 and mAP@(50-95)), F1-score, GFLOPs and inference time. Among all algorithms, the improved YOLOv8 achieved the highest accuracy, with a precision of 92.7%, recall of 91%, mAP@50 of 94.2%, mAP@(50-95) of 67.9%, and F1-score of 92%. With the inclusion of the Focus, Ghost Convolution and C3TR modules, the improved algorithm provided approximately a 1% increase in mAP@50 and F1-score. In terms of computational efficiency, YOLOv5n achieved the optimal GFLOPs value of 4.1, while YOLOv11n achieved the optimal inference time of 1.3 ms.