Wheat Ear Recognition Based on RetinaNet and Transfer Learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate wheat ear counting is crucial for yield estimation but is labor-intensive and costly.
• Wheat ear characteristics, like color, often blend with the background, complicating manual or automated recognition.

Purpose of the Study:

• To evaluate the performance of Faster R-CNN and RetinaNet for automated wheat ear counting.
• To determine the most effective deep learning model for wheat ear recognition across different growth stages and conditions.

Main Methods:

• Comparative analysis of Faster R-CNN and RetinaNet deep learning models.
• Model training and validation using the Global WHEAT dataset and collected image data.
• Transfer learning applied to assess model robustness on diverse datasets.

Main Results:

• RetinaNet achieved higher average accuracy (0.82) than Faster R-CNN (0.72) on the Global WHEAT dataset.
• After transfer learning, RetinaNet showed a higher R² (0.9722) compared to Faster R-CNN (0.8702) on collected data.
• The models demonstrated robustness, with R² values above 0.90 for wheat ears in filling and maturity stages.

Conclusions:

• RetinaNet is a more accurate and robust method for automated wheat ear recognition.
• This study provides a foundation for automated wheat ear counting and improved yield prediction systems.