Pest-YOLO: A model for large-scale multi-class dense and tiny pest detection and counting.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Frequent agricultural pest outbreaks severely impact crop production.
• Automatic pest monitoring and precise recognition are crucial for agricultural planting.
• Deep learning methods have advanced pest detection, but field application challenges remain.

Purpose of the Study:

• To develop an effective pest detector for multi-category, dense, and tiny pests in field environments.
• To improve the accuracy and recall of pest detection, especially for challenging samples.

Main Methods:

• Introduced a modified loss function incorporating focal loss with weight distribution to enhance attention to hard samples.
• Implemented a non-Intersection over Union bounding box selection and suppression algorithm called the confluence strategy to handle occlusion and adhesion.
• Validated the Pest-YOLO model on the large-scale Pest24 dataset.

Main Results:

• Pest-YOLO achieved 69.59% mAP and 77.71% mRecall on the 24-class Pest24 dataset.
• Demonstrated significant improvements over the benchmark YOLOv4 model (5.32% mAP, 28.12% mRecall).
• Outperformed several other state-of-the-art detection methods.

Conclusions:

• Pest-YOLO effectively addresses challenges in detecting dense and tiny agricultural pests in field settings.
• The model offers a practical solution for enhancing agricultural pest monitoring and management.
• The proposed methods contribute to advancing deep learning applications in precision agriculture.