DC-YOLO: an improved field plant detection algorithm based on YOLOv7-tiny.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Automated weeding is crucial for efficient agricultural production.
• Accurate plant detection is a key challenge, especially distinguishing similar-looking crops and weeds.
• Existing lightweight models struggle with the visual similarities between corn seedlings and weeds.

Purpose of the Study:

• To develop an improved, lightweight object detection model for accurate and efficient plant detection in agriculture.
• To address the challenges posed by the visual similarities between corn and weeds.
• To enhance feature extraction and representation for better detection performance.

Main Methods:

• Proposed an improved YOLOv7-tiny model named DC-YOLO.
• Introduced a Dual Coordinate Attention (DCA) model to enhance feature extraction.
• Integrated the Content-Aware ReAssembly of FEatures (CARAFE) operator for learnable feature reorganization.
• Decoupled the detection head to minimize feature conflicts.

Main Results:

• Achieved 95.7% mean Average Precision (mAP@0.5) on corn and weed datasets.
• Demonstrated a computational effort of 13.083 Giga Floating-point Operations (GFLOPs).
• Maintained a parameter size of 5.223 Million (M).
• Outperformed other mainstream lightweight target detection models.

Conclusions:

• DC-YOLO offers a superior solution for automated weeding through enhanced plant detection.
• The model provides a strong balance of high accuracy and computational efficiency.
• This approach represents a significant advancement in agricultural robotics and precision farming.