An improved algorithm based on YOLOv5 for detecting Ambrosia trifida in UAV images

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Detecting weeds like Ambrosia trifida in UAV images is challenging due to small size, high density, and overlapping leaves.
• Existing object detection algorithms often struggle with these complex field conditions, leading to low accuracy.

Purpose of the Study:

• To develop an improved unmanned aerial vehicle (UAV) image detection algorithm for accurate Ambrosia trifida recognition.
• To enhance the detection accuracy for small, dense, and overlapping leafy weed targets.

Main Methods:

• Proposed the YOLOv5-KE algorithm, an enhancement of YOLOv5.
• Incorporated a micro-scale detection layer and adjusted hierarchical detection settings using k-Means for Anchor Box.
• Improved the CIoU loss function and optimized the detection box fusion algorithm.

Main Results:

• Achieved a best detection accuracy of 93.9% for Ambrosia trifida in UAV images.
• Demonstrated a 15.2% improvement over the original YOLOv5 algorithm.
• Outperformed other state-of-the-art algorithms including YOLOv7, YOLOv8, YOLO-NAS, RT-DETR, Faster RCNN, SSD, and Retina Net.

Conclusions:

• YOLOv5-KE is a practical and effective algorithm for detecting Ambrosia trifida in complex field conditions.
• The algorithm shows significant potential for detecting similar small, dense, and overlapping leafy weeds using UAV imagery.
• Provides a valuable technical reference for weed detection applications in precision agriculture.