Broad Dataset and Methods for Counting and Localization of On-Ear Corn Kernels

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop monitoring and yield prediction are crucial for effective farm management.
• Manual corn kernel counting for yield estimation is labor-intensive and impractical due to the high number of kernels per ear.
• Traditional estimation methods lack precision, highlighting the need for automated solutions.

Purpose of the Study:

• To evaluate deep learning frameworks for automated corn kernel counting and localization on corn ears.
• To compare the performance of detection-based methods (Faster R-CNN, YOLO) against density-estimation approaches.
• To introduce a new dataset for advancing research in corn kernel counting and related computer vision tasks.

Main Methods:

• Comparison of Faster R-CNN, YOLOv5, and density-estimation techniques for corn kernel counting.
• Evaluation of model accuracy, edge-deployability, and computational efficiency.
• Development and release of a novel dataset featuring high-quality, multi-class segmentation masks of corn ears.

Main Results:

• YOLOv5 demonstrated accuracy and edge-deployability for corn kernel counting.
• The density-estimation approach achieved high-quality results, was lightweight for edge deployment, and maintained computational efficiency regardless of kernel count.
• A new, challenging dataset was released to facilitate quantitative comparisons and further research.

Conclusions:

• Deep learning models, particularly YOLOv5 and density-estimation, provide accurate and efficient solutions for corn kernel counting.
• The developed dataset will serve as a benchmark for future research in agricultural computer vision, including transfer learning and edge deployment.
• Automated kernel counting offers a significant improvement over manual methods for precision agriculture.