YOLO-CFruit: a robust object detection method for Camellia oleifera fruit in complex environments

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Automated harvesting of Camellia oleifera fruit is crucial for agricultural efficiency.
• Accurate fruit detection in natural environments is challenging due to factors like shadows.
• Traditional methods struggle with variable lighting and complex backgrounds.

Purpose of the Study:

• To develop an efficient deep learning method for accurate Camellia oleifera fruit detection.
• To address the limitations of existing methods in challenging natural environments.
• To enhance the performance of automated harvesting systems.

Main Methods:

• Proposed YOLO-CFruit, a deep learning model integrating CBAM and CSP with Transformer modules.
• Created and enhanced a dataset of Camellia oleifera fruit images.
• Replaced CIoU Loss with EIoU Loss in the YOLOv5 architecture.

Main Results:

• Achieved 98.2% average precision, 94.5% recall, and 98% accuracy.
• Demonstrated a 1.2% improvement in average precision over conventional YOLOv5s.
• Obtained a high F1 score of 96.2 and a processing speed of 19.02 ms.

Conclusions:

• YOLO-CFruit exhibits robust performance in diverse real-world conditions, including varying light and shadow.
• The model's high reliability supports the development of automated Camellia oleifera fruit picking devices.
• This research advances precision agriculture through improved fruit detection technology.