A YOLO11-MSC-Based Method for Small Pest Detection in Insect Pest Monitoring Lamp Images

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Insect pest monitoring faces challenges with small target detection, dense populations, and background interference.
• Existing models often struggle with missed detections and low accuracy in complex agricultural settings.

Purpose of the Study:

• To develop a lightweight and accurate agricultural pest detection model for insect monitoring lamps.
• To improve the detection of small and densely distributed pests in challenging visual conditions.

Main Methods:

• Proposed YOLO11-MSC model featuring Object-Centric Adaptive Slicing (OCAS) dataset construction.
• Introduced Multi-path Nonlinear Star Aggregation (MNS) and C2-based Improved Bi-Level Routing Attention (C2iBRA) modules for enhanced feature extraction.
• Developed Semantic-Guided Adaptive Slicing (SGAS) for efficient full-image inference.

Main Results:

• YOLO11-MSC achieved 95.2% mAP@0.5, 91.8% Precision, and 93.0% Recall on the Insect35_OCAS dataset.
• The model boasts low complexity with 3.4M parameters and 7.8 GFLOPs.
• With SGAS, YOLO11-MSC reached 86.4% mAP@0.5 on original high-resolution images.

Conclusions:

• The YOLO11-MSC model effectively enhances small pest detection accuracy and full-image inference capabilities.
• The proposed methods provide technical support for intelligent agricultural pest monitoring and precision control.
• The model maintains low complexity, making it suitable for practical agricultural applications.