OE-YOLO: An EfficientNet-Based YOLO Network for Rice Panicle Detection.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate rice panicle detection is crucial for precision agriculture but challenging due to environmental complexities.
• Existing methods struggle with small, dense, and variably oriented rice panicles, often confusing them with background elements.

Purpose of the Study:

• To develop an enhanced deep learning framework, OE-YOLO, for accurate and efficient rice panicle detection in complex field conditions.
• To improve upon existing object detection models by incorporating oriented bounding boxes and advanced network modules.

Main Methods:

• OE-YOLO utilizes oriented bounding boxes (OBB) instead of horizontal bounding boxes (HBB) for precise feature capture.
• The backbone network is enhanced with EfficientNetV2 for balanced multi-scale feature extraction and computational efficiency.
• A dynamic convolution-based C3k2_DConv module is employed to amplify discriminative features and reduce background interference.

Main Results:

• OE-YOLO achieved 86.9% mAP50 on rice UAV imagery, outperforming YOLOv8-obb and YOLOv11 by 2.8% and 8.3%, respectively.
• The model demonstrates strong generalization across different flight heights (3m, 10m) and growth stages (heading, filling).
• OE-YOLO offers a computationally frugal solution with only 2.45 M parameters and 4.8 GFLOPs.

Conclusions:

• OE-YOLO provides a highly accurate and computationally efficient solution for real-time rice panicle detection.
• The proposed framework addresses critical needs in precision agriculture for robust, oriented detection under resource constraints.
• This advancement supports improved crop monitoring and management through advanced computer vision techniques.