A Lightweight Pig Aggressive Behavior Recognition Model by Effective Integration of Spatio-Temporal Features

Area of Science:

• Agricultural technology
• Computer vision
• Animal behavior analysis

Background:

• Smart agriculture and pig farming expansion necessitate efficient methods for monitoring herd health.
• Environmental variations like lighting and background in barns challenge existing pig aggressive behavior recognition systems, leading to detection errors.

Purpose of the Study:

• To develop an adaptable deep learning model for pig aggressive behavior recognition in complex farming environments.
• To enhance the accuracy and efficiency of automated detection of aggressive behaviors in pigs.

Main Methods:

• A novel model integrating MobileNetV2 and Autoformer for feature extraction and temporal analysis.
• Incorporation of Convolutional Block Attention Module (CBAM) and Advanced Filtering Feature Fusion Pyramid Network (HS-FPN) within MobileNetV2 for improved feature capture and small target detection.
• Utilizing an improved Autoformer with Gate Attention Unit (GAU) for efficient temporal correlation analysis and focus on critical features.

Main Results:

• The proposed model achieved high performance metrics: 98.08% recall, 94.44% precision, 96.23% accuracy, and 96.23% F1-score.
• Optimized parameter quantity to 10.41 M, demonstrating computational efficiency.
• Outperformed existing models like MobileNetV2-LSTM and MobileNetV2-GRU in accuracy by 3.5% and 3.0%, respectively.

Conclusions:

• The developed model offers a robust balance between recognition accuracy and computational complexity for practical pig farming.
• This technology provides essential data support for scientific feeding and management strategies in automated pig farming.
• The model's adaptability to complex environments makes it suitable for real-world smart agriculture applications.