DPD-YOLO: dense pineapple fruit target detection algorithm in complex environments based on YOLOv8 combined with...

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Drones and computer vision are increasingly used for agricultural yield estimation.
• Detecting pineapples with drones is challenging due to fruit occlusion by leaves and complex field backgrounds.
• Existing target detection algorithms struggle with small, occluded targets in complex environments.

Purpose of the Study:

• To develop an improved deep learning model for accurate pineapple detection in challenging agricultural settings.
• To enhance the performance of drone-based pineapple yield estimation systems.

Main Methods:

• Proposed DPD-YOLO (Dense-Pineapple-Detection YOU Only Look Once) model based on YOLOv8.
• Integrated Coordinate Attention mechanism for feature extraction and Bi-directional Feature Pyramid Network (BiFPN) for multi-scale feature fusion.
• Replaced YOLOv8 detection head with RT-DETR (Real-Time Detection Transformer) incorporating attention mechanisms and utilized Focaler-IoU for improved small target detection.

Main Results:

• DPD-YOLO demonstrated superior performance compared to mainstream models in complex, occluded pineapple fields.
• Achieved a mean Average Precision (mAP@0.5) of 62.0%, a 6.6% improvement over YOLOv8.
• Showcased significant improvements in Precision (2.7%), Recall (13%), and F1-score (10.3%).

Conclusions:

• The DPD-YOLO model effectively addresses the limitations of current algorithms for detecting occluded pineapples in complex backgrounds.
• This enhanced model offers a promising solution for accurate drone-based pineapple yield estimation in precision agriculture.