A scalable multi-modal learning fruit detection algorithm for dynamic environments.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Detecting litchi fruits in natural scenes is challenging due to dense occlusion and small target sizes.
• Existing methods struggle with accurate identification and localization in complex agricultural environments.

Purpose of the Study:

• To develop a novel multimodal target detection method, YOLOv5-Litchi, for enhanced litchi fruit detection.
• To address the limitations of dense occlusion and small target identification in litchi fruit detection.

Main Methods:

• Modified YOLOv5s architecture with a simplified FPN structure and increased detection heads (5).
• Incorporated TSCD detection heads for small targets and replaced loss functions with EIoU and VFLoss.
• Utilized a sliding slice method for improved small target prediction.

Main Results:

• YOLOv5-Litchi achieved improvements of 9.5% in accuracy, 0.9% in recall, and 12.3% in mean average precision (mAP) over the original YOLOv5s.
• Outperformed YOLOx, YOLOv6, and YOLOv8 with higher AP values (4.0%, 6.3%, and 3.7% respectively).

Conclusions:

• The enhanced network significantly improves recall and AP, reducing missed detections and enhancing prediction frame accuracy.
• YOLOv5-Litchi effectively handles dense occlusion and small targets, proving suitable for litchi fruit detection and yield estimation.