MPE-YOLO: an efficient and lightweight model for maize plug-tray empty-cell detection and iOS deployment

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate seedling quality monitoring is crucial for large-scale greenhouse production.
• Manual inspection of maize plug trays is labor-intensive and subjective.
• Automated empty-cell recognition faces challenges like occlusion and substrate interference.

Purpose of the Study:

• To develop a lightweight and efficient model for empty-cell recognition in maize plug trays.
• To address limitations of existing methods in complex greenhouse environments.
• To enable portable deployment for real-time seedling quality assessment.

Main Methods:

• Proposed MPE-YOLO, a lightweight YOLOv26n-based model incorporating depthwise separable convolution and CGSB modules.
• Introduced an EMA-enhanced C3k2 module (C3k2-EMA) for improved attention to empty cells.
• Developed a dual-focusing loss function (Focal Loss + Wise-IoU v3) for enhanced learning.
• Deployed the optimized model on an iPhone for mobile application.

Main Results:

• MPE-YOLO improved precision by 1.6% and mAP@0.5 by 0.3% over the original YOLOv26n, with a 16.0% increase in inference speed.
• Demonstrated superior robustness in complex backgrounds with occlusion and substrate coverage.
• Significantly reduced false detections per tray from 3.1 to 1.3.
• Validated the feasibility of portable detection using an iPhone-based application.

Conclusions:

• MPE-YOLO offers an efficient and portable solution for maize plug-tray empty-cell recognition.
• The developed model enhances seedling quality monitoring and automated transplanting operations.
• The mobile application facilitates practical, on-site deployment for greenhouse management.