YOLOv5s-FP: A Novel Method for In-Field Pear Detection Using a Transformer Encoder and Multi-Scale Collaboration...

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Orchard management modernization relies on precise pear detection.
• Challenges include small, occluded pears, leading to high false detection and object loss rates.

Purpose of the Study:

• To propose a multi-scale collaborative perception network (YOLOv5s-FP) for improved pear detection.
• To address challenges of small and occluded pears in diverse orchard environments.

Main Methods:

• Developed a pear dataset with 3680 images from ground and UAV platforms.
• Optimized the cross-stage partial (CSP) module with a transformer encoder for global feature extraction.
• Integrated an attentional feature fusion mechanism and a modified path aggregation network for multi-scale feature collaboration.

Main Results:

• Achieved the highest average precision (96.12%) compared to other YOLO series networks.
• Demonstrated superior robustness to occlusion and illumination variations.
• Showcased effective detection of various pear sizes in dense, overlapping, and low-light conditions.

Conclusions:

• The YOLOv5s-FP network offers a practical, real-time, and accurate solution for in-field pear detection.
• It supports advanced pear growth monitoring and automated harvesting in unmanned orchards.