EBE-YOLOv4: A lightweight detecting model for pine cones in forest.

Area of Science:

• Computer Vision
• Machine Learning
• Forestry Technology

Background:

• Pine cones are valuable forest products, but their collection is challenging due to dispersed distribution and complex picking processes.
• Existing object detection models often have high computational costs, limiting their application in real-time forest monitoring.

Purpose of the Study:

• To develop a lightweight and accurate machine vision model for rapid pine cone recognition in forest environments.
• To address the limitations of general YOLOv4 models in terms of parameter count and computational efficiency.

Main Methods:

• A novel EBE-YOLOv4 model was designed, utilizing EfficientNet-b0 as the backbone and incorporating a channel-transformed BiFPN structure with ECA-Net attention in the neck.
• The H-Swish activation function was employed to optimize model performance and accuracy.
• Experimental data comprised 768 pine cone images, expanded to 1536, and divided into training (80%) and testing (20%) sets.

Main Results:

• The EBE-YOLOv4 model achieved a precision of 96.25% and a recall rate of 82.72%.
• The model demonstrated a significant increase in detection speed, reaching 64.09 FPS, a 70% improvement over the original YOLOv4.
• The lightweight design maintained detection accuracy while substantially enhancing processing speed.

Conclusions:

• The EBE-YOLOv4 model offers an effective and efficient solution for automated pine cone detection in forests.
• The optimized architecture successfully balances accuracy and computational performance, paving the way for practical applications in forest resource management.