APO-CViT: A Non-Destructive Estrus Detection Method for Breeding Pigs Based on Multimodal Feature Fusion.

Area of Science:

• Animal Science
• Artificial Intelligence
• Biotechnology

Background:

• Traditional sow estrus detection methods are subjective and inaccurate, hindering efficient pig reproductive performance.
• Modern swine farming demands objective, reliable, and non-destructive estrus monitoring techniques.
• Existing methods fail to meet the accuracy requirements for optimizing breeding programs.

Purpose of the Study:

• To develop a multimodal feature fusion method for accurate and robust estrus detection in sows.
• To enhance pig reproductive performance and farm production efficiency through improved estrus monitoring.
• To create an objective, non-destructive system for identifying estrus states in breeding pigs.

Main Methods:

• Developed the Adaptive-PIG-OESTUS-CNN-ViT model, integrating audio and thermal infrared image data.
• Employed a fusion strategy combining Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) for feature extraction.
• Utilized an adaptive cross-attention mechanism for multimodal feature vector learning and an improved DenseNet for classification.

Main Results:

• The multimodal model achieved high performance metrics: 98.92% accuracy, 95.83% recall, and 97.35% F1-score.
• Demonstrated effective non-destructive estrus detection in breeding pigs.
• The model accurately distinguished between estrus and non-estrus states using fused audio and thermal data.

Conclusions:

• The Adaptive-PIG-OESTUS-CNN-ViT model offers a significant advancement in sow estrus detection.
• This multimodal approach provides an efficient, objective, and non-destructive solution for modern swine farming.
• The findings support the integration of AI and multimodal data for optimizing animal reproductive management.