SP-YOLO: A Real-Time and Efficient Multi-Scale Model for Pest Detection in Sugar Beet Fields

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Beet crops face significant yield losses due to pest infestations.
• Accurate and timely pest identification is critical for effective crop management.
• Current pest detection methods struggle with camouflaged pests and scale variations.

Purpose of the Study:

• To develop an improved real-time pest detection model for beet crops.
• To address challenges in pest detection, including environmental blending and scale variations.
• To construct a multi-scale pest dataset for complex backgrounds.

Main Methods:

• Developed the BeetPest dataset for multi-scale pest detection in complex backgrounds.
• Proposed the SP-YOLO model, an enhanced real-time detection model based on YOLOv11.
• Integrated a CNN and transformer (CAT) into the backbone for global feature capture.
• Utilized a lightweight depthwise separable convolution block (DSCB) for multi-scale feature extraction.
• Employed a cross-layer path aggregation network (CLPAN) in the neck for feature merging.

Main Results:

• SP-YOLO demonstrated a 4.9% improvement in mean average precision (mAP@50) over YOLOv11.
• Achieved a 9.9% increase in precision and a 1.3% rise in average recall.
• Reached a detection speed of 136 frames per second (FPS), enabling real-time application.
• Showcased robustness on other pest datasets with manageable computational complexity.

Conclusions:

• The SP-YOLO model effectively detects pests in complex backgrounds and handles scale variations.
• The model meets real-time detection requirements for agricultural pest management.
• SP-YOLO is suitable for edge devices due to its efficiency and manageable parameter size.