GIWT-YOLO: an efficient multi-scale framework for real-time Scolytinae pests detection.

Area of Science:

• Agricultural Entomology
• Computer Vision
• Machine Learning

Background:

• Scolytinae pests pose significant challenges to agriculture due to their size variation and visual similarities.
• Mainstream object detection models struggle with accurate identification, particularly for smaller pest species.

Purpose of the Study:

• To develop a lightweight and accurate object detection model for Scolytinae pests.
• To improve the detection of small and visually similar pest species.

Main Methods:

• Proposed GIWT-YOLO, a lightweight model based on YOLOv11s, incorporating a GIConv module for multi-scale feature extraction and a WTConv module inspired by wavelet transform.
• Integrated an SE attention mechanism to enhance focus on key feature regions.
• Evaluated model performance on Scolytinae pest detection datasets.

Main Results:

• GIWT-YOLO achieved 84.7% Precision, 88.7% mAP@50, and 63.4% mAP@50~95, outperforming the baseline YOLOv11s.
• Reduced model parameters by 11.3% and GFLOPs by 13.4%, indicating a more efficient architecture.
• Demonstrated state-of-the-art performance in small-sized pest detection with strong generalization ability.

Conclusions:

• GIWT-YOLO offers an efficient and accurate solution for detecting Scolytinae pests, especially small and similar-looking species.
• The model's lightweight design makes it suitable for practical applications in pest management.
• Future work includes expanding the dataset to enhance applicability across a wider range of pest scenarios.