A Lightweight model for accurate soybean pod detection: YOLO-MobilePod

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Soybean pod detection is crucial for yield and quality assessment.
• Natural environments present challenges like small size, dense distribution, and occlusion.
• Existing object detection models struggle with these small-scale, overlapping targets.

Purpose of the Study:

• To develop a lightweight and efficient soybean pod detection model.
• To enhance detection accuracy for small and overlapping soybean pods.
• To provide a feasible technical reference for efficient soybean pod detection.

Main Methods:

• Developed YOLO-MobilePod based on YOLOv12, using MobileNetV4 as the backbone.
• Implemented high-resolution feature fusion in the neck network for improved small-object feature representation.
• Incorporated dynamic convolution (DynamicConv) in the detection head for enhanced feature modeling and lightweighting.

Main Results:

• YOLO-MobilePod achieved a 2.07% increase in recall and a 4.10% increase in mAP50-95 compared to YOLOv12n.
• Reduced parameters by 45.31%, FLOPs by 4.762%, and model size by 38.18%.
• Increased inference speed by 8.768%.

Conclusions:

• The proposed YOLO-MobilePod model effectively recognizes diverse soybean pods.
• It achieves improved detection accuracy while maintaining a lightweight design.
• Offers a practical solution for efficient soybean pod detection in agricultural applications.