An improved YOLO v4 used for grape detection in unstructured environment

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual recognition is crucial for harvesting robot accuracy.
• Unstructured environments (occlusion, lighting changes, fog) challenge current detection algorithms.

Purpose of the Study:

• To propose an improved YOLO v4 model (YOLO v4+) for robust fruit detection in unstructured environments.
• To enhance feature extraction and reduce information loss for better accuracy.

Main Methods:

• Implemented a parameterless attention mechanism in the backbone.
• Introduced a multi-scale feature fusion module with fusion weight and jump connections.
• Utilized focal loss with adjusted hyperparameters (α=0.75, γ=2).

Main Results:

• YOLO v4+ achieved 94.25% average precision and 93% F1 score.
• Demonstrated a 3.35% increase in average precision and 3% in F1 score over YOLO v4.
• Outperformed other state-of-the-art models in comprehensive and generalization ability.

Conclusions:

• The proposed YOLO v4+ model significantly enhances detection accuracy and robustness for harvesting robots.
• Tailored augmentation methods further improve model performance in specific working conditions.
• The method holds potential to increase the applicability and robustness of robotic harvesting systems.