Oriented feature pyramid network for small and dense wheat heads detection and counting.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate wheat head detection is crucial for precision agriculture applications like yield estimation.
• Challenges include variations in size, orientation, density, and imaging conditions, especially for small and dense wheat heads.

Purpose of the Study:

• To propose a novel deep learning approach, the Oriented Feature Pyramid Network (OFPN), for accurate wheat head detection and counting.
• To introduce a new dataset, the Rotated Global Wheat Head Dataset (RGWHD), for evaluating oriented wheat head detection.

Main Methods:

• Developed the Oriented Feature Pyramid Network (OFPN) using oriented bounding boxes for rotated wheat head detection.
• Integrated Path-aggregation and Balanced Feature Pyramid Network for multi-scale feature fusion.
• Utilized Soft-NMS algorithm for improved localization of dense and overlapping wheat heads.
• Created the Rotated Global Wheat Head Dataset (RGWHD) with manually annotated oriented bounding boxes.

Main Results:

• OFPN achieved a mean average precision of 85.77% for oriented wheat head detection, outperforming six state-of-the-art models.
• Wheat head counting accuracy reached 93.97%, a 3.12% improvement over Faster R-CNN.
• Demonstrated superior performance in localizing and counting wheat heads in challenging scenarios.

Conclusions:

• The OFPN model effectively addresses the challenges of detecting and counting small, dense, and rotated wheat heads.
• OFPN shows significant potential for enhancing precision agriculture through improved wheat monitoring and yield prediction.
• The RGWHD dataset provides a valuable resource for future research in wheat head detection.