Leaf Counting: Fusing Network Components for Improved Accuracy.

Area of Science:

• Computer Vision
• Plant Phenotyping
• Deep Learning

Background:

• Accurate leaf counting is crucial for assessing plant health and growth.
• Visual phenotyping increasingly relies on automated leaf counting techniques.
• Existing methods face challenges with varying leaf scales and complex plant structures.

Purpose of the Study:

• To develop and evaluate novel deep learning approaches for precise visual leaf counting.
• To compare the performance of direct regression and detection-based counting methods.
• To establish a new state-of-the-art in automated leaf counting for plant phenotyping.

Main Methods:

• A multi-scale regression approach using fused image resolutions to count leaves.
• A detection-based method that locates leaf centers and aggregates counts.
• Evaluation on the Leaf Counting Challenge (LCC) dataset and a new banana leaf dataset.

Main Results:

• Both proposed deep learning methods surpassed previous state-of-the-art results on the LCC dataset.
• The detection-based method achieved 0.94 average precision for leaf center localization.
• The multi-scale regression method proved effective when center-dot annotations were unavailable.

Conclusions:

• The developed deep learning models significantly advance automated leaf counting capabilities.
• The choice between methods depends on dataset characteristics and annotation availability.
• These findings offer robust solutions for plant health monitoring and phenotyping research.