YOLOv8-BVC: Lightweight Industrial Defect Detection

Currently, in the field of industrial surface defect detection, generic object detection algorithms are complex and computationally intensive. However, many small and mediumsized enterprises responsible for inspections often have devices with limited computational capabilities and resources, making it challenging to deploy these detection algorithms. To address this issue, this paper proposes a lightweight steel surface defect detection model, YOLOv8-BVC, based on the YOLOv8n object detection framework.YOLOv8-BVC incorporates a lightweight VanillaNet in the backbone feature extraction network to reduce unnecessary branch structures, thereby lowering the model’s complexity. Additionally, a BOT module, combining the strengths of CNN and Transformer, is embedded after the VanillaNet module to enhance feature information processing and improve detection speed. This allows the model to achieve efficient inference speeds even on handheld devices with low computational capabilities. Furthermore, CA modules are integrated at the end of the YOLOv8n backbone and after the C2f neck to enhance feature extraction capabilities, emphasizing critical information for surface crack detection, thus improving accuracy and model generalization across various scenarios.BVC-YOLO achieves the best mAP of 97.4% during training. Although it has a slightly lower mAP compared to segnet (a decrease of 2.5%), BVC-YOLO significantly reduces model size by nearly 60MB, with only 1. 96M parameters and 6.0 GFLOPs, while achieving an FPS of 263/s. These results demonstrate that BVC-YOLO is suitable for deployment on handheld devices.