An optimized YOLOv8n based model for real time defect detection in taro strip production

Taro production is predominantly manual, creating a need for automated defect detection in processing lines to improve efficiency and product quality. Traditional computer vision methods lack accuracy under practical industrial production conditions, while deep learning approaches, though robust, are often computationally expensive. This study presents a modified YOLOv8n model designed for automated defect detection of taro strips in industrial production environments where real-time processing and high accuracy are critical. The proposed model integrates several architectural innovations: a Bi-directional Feature Pyramid Network (BiFPN) to facilitate improved multi-scale feature fusion, the VoV-GSCSP module to replace the conventional C2f block and reduce computational complexity, and a shared parameter detection head to further lighten the model. In addition, an embedded Wise Intersection over Union (WIoU) loss function is used to accelerate convergence and improve prediction accuracy by optimizing bounding box alignment with ground truth data. Multiple data augmentation strategies, such as noise, motion, lighting, and background, are used in training to mitigate the risk of overfitting and enhance model scalability. Both offline and online experimental results demonstrate that the optimized model achieves an average mean detection accuracy (mAP50) of over 99%, with a precision and recall of over 0.99, and 3.23 G FLOPs. These results are significantly better than those of the original YOLOv8n, which achieved an mAP50 of 94.63%, a precision of 0.955, and a recall of 0.934. When deployed on a Raspberry Pi 5, the modified model demonstrates good robustness and generalization by accurately detecting defects in previously unseen taro-strip data. These results clearly show the advantages of the proposed method in terms of accuracy and computational efficiency, making it ideal for real-time taro-strip defect detection.