Yolo-Pest: An Insect Pest Object Detection Algorithm via CAC3 Module.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Insect pests pose a significant threat to crop yield and quality.
• Existing pest detection algorithms struggle with small targets due to limited data and models.
• Accurate and timely pest detection is crucial for effective agricultural pest control.

Purpose of the Study:

• To improve convolutional neural network (CNN) models for detecting small agricultural pests.
• To propose a lightweight and effective pest detection method, Yolo-Pest, for small target pests.
• To address the challenge of feature extraction in small sample learning for pest detection.

Main Methods:

• Developed the CAC3 module, utilizing a stacking residual structure based on the standard BottleNeck module, to enhance feature extraction for small sample learning.
• Integrated a ConvNext module, inspired by the vision transformer (ViT), to achieve effective feature extraction while maintaining a lightweight network architecture.
• Trained and evaluated the Yolo-Pest model on the Teddy Cup pest dataset and public datasets like IP102.

Main Results:

• The proposed Yolo-Pest method achieved 91.9% mAP0.5 on the Teddy Cup pest dataset, outperforming the Yolov5s model by nearly 8%.
• Demonstrated effective feature extraction and maintained a lightweight network design.
• Showcased strong performance on public datasets, such as IP102, with a significant reduction in parameters.

Conclusions:

• The Yolo-Pest method offers a lightweight and effective solution for agricultural pest detection, particularly for small targets.
• The proposed CAC3 module and ConvNext integration enhance feature extraction capabilities in CNN models for pest detection.
• The approach shows significant improvements over existing methods and holds promise for practical agricultural applications.