Non-destructive detection of micro-impurities in tea using the YOLOv11-PFT model.

Microscopic impurities can contaminate tea during production, processing, and packaging. Current technologies remove only visible contaminants, leaving microscopic foreign objects that compromise tea quality, and reliable detection methods remain lacking. To address this challenge, we propose YOLOv11-PFT, an improved deep learning model based on YOLOv11, enhanced with Powerful-IoU loss, FasterNet, and Triple Attention modules to boost detection accuracy, reduce model size, and improve feature extraction. The resulting lightweight model achieves 99.16% detection accuracy for microscopic tea contaminants, with Precision, Recall, F1 score, and mAP all near 98.7-99.2%, GFLOPs of 5.5, inference speed of 340.6 FPS, and a model size of only 5.0 MB. It outperforms seven benchmark models in accuracy. YOLOv11-PFT offers an effective solution for microscopic contaminant detection in tea, supporting automation in food safety, intelligent quality control, and edge-device deployment in agriculture.