MAR-YOLOv9: A multi-dataset object detection method for agricultural fields based on YOLOv9

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Deep learning object detection models face performance degradation in cross-dataset scenarios, especially in complex agricultural environments.
• Existing technologies struggle with diverse crop types and variable conditions, leading to performance bottlenecks.

Purpose of the Study:

• To propose a lightweight, cross-dataset enhanced object detection method for agriculture using YOLOv9.
• To address performance degradation and computational complexity in agricultural object detection.

Main Methods:

• Developed Multi-Adapt Recognition-YOLOv9 (MAR-YOLOv9) based on YOLOv9.
• Optimized the Backbone network with 16x downsampling and introduced a streamlined Main Neck structure.
• Implemented a hybrid connection strategy for flexible feature utilization.

Main Results:

• MAR-YOLOv9 achieved a 39.18% mAP@0.5 improvement over seven mainstream algorithms and 1.28% over YOLOv9.
• Reduced model size by 9.3% and decreased the number of layers, lowering computational costs.
• Demonstrated significant advantages in detecting complex agricultural images.

Conclusions:

• MAR-YOLOv9 offers an efficient, lightweight, and adaptable solution for real-time object detection in agriculture.
• The model maintains high performance while improving detection speed and reducing resource requirements.
• Provides a valuable tool for addressing challenges in agricultural image analysis.