Real-Time Plant Leaf Counting Using Deep Object Detection Networks

Area of Science:

• Agricultural Informatics
• Computer Vision
• Plant Science

Background:

• Deep neural networks (DNNs) offer insights into plant traits but often process data too slowly for real-time applications.
• Real-time plant phenotyping is essential for precision agriculture and agricultural informatics.

Purpose of the Study:

• To develop and evaluate a real-time system for accurate plant leaf detection, counting, and localization.
• To create and share a new annotated dataset for plant leaf analysis.

Main Methods:

• Utilized state-of-the-art object detection networks, specifically Tiny-YOLOv3, for real-time leaf analysis.
• Created and publicly released an annotated dataset of *Arabidopsis* plants.
• Implemented a robotics platform for demonstrating real-time greenhouse capabilities.
• Compared Tiny-YOLOv3 performance against Faster R-CNN.

Main Results:

• Tiny-YOLOv3 achieved real-time leaf localization and counting with an inference time under 0.01 s.
• Achieved an F1 Score over 0.94 and a False Positive Rate (FPR) around 24% with Tiny-YOLOv3.
• Tiny-YOLOv3 demonstrated faster inference and improved F1 Score compared to Faster R-CNN, though with a higher difference in count (DiC) and lower AP.

Conclusions:

• The developed Tiny-YOLOv3 network enables efficient real-time leaf counting and localization.
• The publicly available dataset and trained network contribute to advancing plant phenotyping research.
• Real-time DNNs show significant potential for enhancing precision agriculture applications.