Leaf-Counting in Monocot Plants Using Deep Regression Models.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate leaf counting is crucial for crop yield estimation.
• Manual leaf counting is labor-intensive and costly.
• Existing deep learning methods struggle with monocot plants and occluded leaves, requiring extensive data.

Purpose of the Study:

• To develop an effective deep neural network for leaf counting in monocot plants (e.g., maize, sorghum).
• To reduce the need for extensive data annotation and labeling.
• To improve leaf counting accuracy, especially in the presence of leaf overlaps and occlusions.

Main Methods:

• A novel deep neural network approach utilizing leaf skeletons and image augmentation.
• Extraction of leaf skeletons to capture topological information.
• Employing a regression model transferred from Inception-Resnet-V2, integrating original images, skeletons, and augmentations.

Main Results:

• The proposed method achieves superior performance in leaf counting for monocot plants, outperforming existing approaches.
• Demonstrated effectiveness even with severe leaf occlusions and overlaps.
• Significantly reduced training costs due to minimal annotation requirements.

Conclusions:

• The developed deep neural network offers an accurate and cost-effective solution for leaf counting in crops like maize.
• The method's robustness against noise and occlusions makes it suitable for real-world agricultural applications.
• This approach lowers the barrier for applying advanced computer vision techniques in crop monitoring.