PD-YOLO: a novel weed detection method based on multi-scale feature fusion

Area of Science:

• Agricultural technology
• Computer vision
• Robotics

Background:

• Automated weeding using robots is key for sustainable agriculture and reduced labor.
• Accurate weed identification via computer vision faces challenges like crop-weed similarity, scale variation, occlusion, and small object sizes.

Purpose of the Study:

• To develop a novel object detection model, PD-YOLO, for enhanced weed detection in complex agricultural environments.
• To improve the accuracy and efficiency of automated weed identification systems.

Main Methods:

• Proposed PD-YOLO model based on the YOLOv8n framework.
• Incorporated a Parallel Focusing Feature Pyramid (PF-FPN) with Feature Filtering and Aggregation Module (FFAM) and Hierarchical Adaptive Recalibration Fusion Module (HARFM).
• Utilized a dynamic detection head (Dyhead) for improved detection in complex scenarios.

Main Results:

• PD-YOLO demonstrated superior performance compared to state-of-the-art models on public weed datasets.
• Achieved a 1.7% and 1.8% improvement in mean average precision (mAP) on the CottonWeedDet12 dataset at thresholds of 0.5 and 0.5-0.95, respectively.
• Showcased a modest increase in computational cost.

Conclusions:

• PD-YOLO offers an efficient and accurate solution for automated weed detection.
• The model provides technological advancements for robotic weeding systems in agriculture.
• This research contributes valuable insights into overcoming computer vision challenges in weed identification.