YOLO-P: An efficient method for pear fast detection in complex orchard picking environment.

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Accurate fruit detection is crucial for automated picking robots.
• Complex orchard environments with disordered backgrounds and shading significantly reduce detection accuracy.

Purpose of the Study:

• To propose an effective model, YOLO-P, for rapid and accurate pear detection.
• To improve the performance of fruit detection in challenging, real-world orchard conditions.

Main Methods:

• Modified YOLOv5 backbone using Shuffle block and inverted shuffle block for efficient long-distance feature extraction.
• Integrated Convolutional Block Attention Module (CBAM) to enhance key feature capture.
• Utilized Hard-Swish activation and a weighted confidence loss function for improved detection, especially for small targets.

Main Results:

• YOLO-P achieved an average precision (AP) of 97.6%, outperforming other lightweight networks and improving original YOLOv5s by 1.8%.
• Model volume was reduced by 39.4% (from 13.7MB to 8.3MB).
• Demonstrated high performance in daytime and nighttime detection experiments across various complexities and shading conditions, achieving an F1 score of 96.1% and 32 FPS.

Conclusions:

• The proposed YOLO-P model significantly enhances pear detection accuracy and efficiency in complex orchard environments.
• YOLO-P's performance is sufficient for real-time robotic picking applications.
• The method offers a valuable reference for automated fruit detection in similar unstructured settings.