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

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Taro processing relies heavily on manual labor, necessitating automated defect detection for improved efficiency and quality.
• Traditional computer vision methods struggle with accuracy in industrial settings, while deep learning models can be computationally intensive.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for automated defect detection of taro strips in real-time industrial environments.
• To address the limitations of existing methods by optimizing model architecture and loss functions for improved performance.

Main Methods:

• A modified YOLOv8n architecture incorporating Bi-directional Feature Pyramid Network (BiFPN) for enhanced feature fusion.
• Integration of the VoV-GSCSP module and a shared parameter detection head to reduce computational complexity.
• Utilization of Wise Intersection over Union (WIoU) loss function and extensive data augmentation for improved accuracy and robustness.

Main Results:

• The optimized model achieved a mean average precision (mAP50) exceeding 99%, with precision and recall above 0.99.
• Demonstrated significantly improved performance compared to the original YOLOv8n model (mAP50: 94.63%).
• The model exhibited robustness and generalization when deployed on a Raspberry Pi 5, accurately detecting defects in new data.

Conclusions:

• The proposed modified YOLOv8n model offers superior accuracy and computational efficiency for real-time taro strip defect detection.
• This advancement is well-suited for industrial applications demanding high-throughput quality control.
• The study highlights the potential of optimized deep learning for agricultural product processing.