A Novel Improved YOLOv3-SC Model for Individual Pig Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animal Science

Background:

• Accurate individual pig detection is crucial for intelligent breeding and health monitoring.
• Current manual methods are inefficient and unfeasible for large-scale operations.
• Improved detection enhances pork production, quality, and economic outcomes.

Purpose of the Study:

• To develop an efficient and effective automated model for individual pig detection.
• To enhance the YOLOv3 architecture with attention mechanisms for richer feature extraction.
• To improve the robustness and accuracy of pig identification in farming environments.

Main Methods:

• Development of the YOLOv3-SC model, integrating Spatial Pyramid Pooling (SPP) and Convolutional Block Attention Module (CBAM).
• Training and evaluation on a custom dataset of 4019 pig images.
• Comparative analysis against YOLOv1, YOLOv2, Faster-RCNN, and YOLOv3.

Main Results:

• The YOLOv3-SC model achieved a mean Average Precision (mAP) of 99.24% for individual pig identification.
• Detection time was recorded at 16 ms, demonstrating high efficiency.
• Significant mAP improvements were observed compared to other leading models: 2.31% (YOLOv1), 1.44% (YOLOv2), 1.28% (Faster-RCNN), and 0.61% (YOLOv3).

Conclusions:

• The proposed YOLOv3-SC model offers accurate and rapid individual pig detection.
• This novel model can be effectively deployed for real-time monitoring on farms.
• It provides innovative solutions for automated individual pig identification, advancing intelligent agriculture.