Lightweight Vision-Transformer Network for Early Insect Pest Identification in Greenhouse Agricultural Environments

Area of Science:

• Horticultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Early detection of plant diseases and pests in greenhouses is crucial for crop yield and quality.
• Real-time deployment of detection systems on edge devices faces challenges with computational efficiency and accuracy.
• Existing models struggle with complex environmental conditions and small-scale targets common in greenhouses.

Purpose of the Study:

• To develop a lightweight, efficient, and accurate intelligent recognition network for automated greenhouse pest and disease detection.
• To address the limitations of current models in handling complex environments and small targets.
• To enable real-time pest and disease identification on resource-constrained edge devices.

Main Methods:

• Proposed Light-HortiNet, a lightweight cross-scale intelligent recognition network.
• Utilized a Mobile-Transformer backbone with a cross-scale lightweight attention mechanism.
• Integrated a small-object enhancement branch and an alternative block distillation strategy.

Main Results:

• Achieved high detection performance with mAP@50 of 0.872 and mAP@50:95 of 0.561.
• Demonstrated superior classification accuracy (0.894), precision (0.886), recall (0.879), and F1-score (0.882).
• Showcased enhanced small-object recognition (mAP-small: 0.536, recall-small: 0.589) and real-time inference (>20 FPS on Jetson Nano).

Conclusions:

• Light-HortiNet offers a robust and stable solution for automated greenhouse pest and disease detection.
• The model significantly outperforms mainstream lightweight networks in accuracy and efficiency.
• Favorable deployment adaptability on edge devices makes it suitable for real-time agricultural monitoring.