A real-time object detection model for orchard pests based on improved YOLOv4 algorithm.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate real-time orchard pest detection is crucial for the fruit industry's economic viability.
• Existing methods may lack efficiency and accuracy in complex field conditions.
• A comprehensive dataset, PestImgData, was created for training and validation.

Purpose of the Study:

• To develop and evaluate an efficient deep learning model for real-time orchard pest detection.
• To investigate the impact of transfer learning, activation functions, anchor boxes, and batch normalization on pest detection.
• To improve upon the YOLOv4 algorithm for enhanced pest identification accuracy and speed.

Main Methods:

• A dataset of 24,796 images covering 7 orchard pest types was utilized.
• The YOLOv4 algorithm was studied, incorporating transfer learning, activation functions, anchor boxes, and batch normalization.
• Improvements included upgrading the NMS algorithm to DIoU-NMS, employing 2-time finetuning, and data enhancement.

Main Results:

• The enhanced model, F-D-YOLOv4-PEST, achieved a mean average precision of 92.86%.
• The model demonstrated a detection time of 12.22 ms per image, meeting real-time requirements.
• Effective performance was maintained even with high pest density or overlap, and in diverse environments (lab, greenhouse, wired/wireless networks).

Conclusions:

• The F-D-YOLOv4-PEST model offers a significant advancement in real-time orchard pest detection.
• The study provides a valuable technical reference for intelligent agricultural pest identification using deep learning.
• The developed model can assist in improving the economic benefits of the fruit industry through efficient pest management.