Tomato Pest Recognition Algorithm Based on Improved YOLOv4.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Tomato production faces significant losses due to diseases and insect pests.
• Current manual pest detection is labor-intensive, costly, and time-consuming.
• Existing automated methods lack accuracy and struggle with complex backgrounds.

Purpose of the Study:

• To develop an advanced deep learning algorithm for accurate tomato pest identification.
• To overcome limitations of existing manual and automated pest detection methods.
• To improve the efficiency and economic benefits for tomato growers.

Main Methods:

• Developed an improved YOLOv4 algorithm fused with a triplet attention mechanism (YOLOv4-TAM).
• Utilized a focal loss function to address class imbalance in pest datasets.
• Employed K-means++ clustering for optimal anchor box selection.
• Created and utilized a labeled dataset of tomato pests for training and testing.

Main Results:

• The YOLOv4-TAM algorithm achieved an average recognition accuracy of 95.2% on the established dataset.
• The proposed method demonstrated superior performance compared to previous pest detection techniques.
• The algorithm proved effective in detecting tomato pests in practical, real-world image conditions.

Conclusions:

• The YOLOv4-TAM algorithm offers a highly accurate and feasible solution for tomato pest detection.
• This deep learning approach significantly improves upon existing methods, aiding in crop protection.
• The developed method has practical applications for enhancing agricultural productivity and reducing economic losses.