Few-shot object detection for pest insects via features aggregation and contrastive learning

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate pest insect detection is vital for crop yield and agricultural management.
• Traditional methods face challenges with species diversity, individual variations, limited data, small target sizes, and complex environments.

Purpose of the Study:

• To develop a novel few-shot object detection (FSOD) method for improved pest insect detection.
• To address limitations of traditional methods in accuracy and data dependency.

Main Methods:

• Utilized Faster R-CNN framework with multi-scale feature extraction (Feature Pyramid Network).
• Implemented a Feature Aggregation Module (FAM) with attention for fusing contextual features.
• Employed supervised contrastive learning (SCL) to enhance feature discriminability.
• Integrated focal loss and class weights to manage class imbalance and focus on difficult samples.

Main Results:

• The proposed FSOD method significantly outperformed existing approaches (YOLO, TFA, VFA, FSCE) on the PestDet20 dataset.
• Achieved superior mean Average Precision (mAP) in 3-shot, 5-shot, and 10-shot scenarios.
• Ablation studies confirmed the substantial contribution of each method component to performance.

Conclusions:

• The research offers a practical and efficient solution for pest detection, reducing reliance on extensive annotated datasets.
• The method demonstrates robustness and improved accuracy for minority pest classes in challenging conditions.
• Despite higher computational complexity than real-time frameworks, the accuracy gains justify its use in critical pest management.