Enhancing Pig Behavior Recognition in Complex Environments: A Transfer Learning-Assisted YOLO11 Network with Wavelet...

Area of Science:

• Agricultural Technology
• Computer Vision
• Animal Science

Background:

• Accurate pig behavior recognition is crucial for disease detection, animal welfare, and precision agriculture.
• Existing deep learning models are often too complex or resource-intensive for practical deployment in farming environments.
• There is a need for efficient and generalizable models for real-time pig behavior analysis on resource-limited devices.

Purpose of the Study:

• To develop a lightweight and efficient deep learning model for pig behavior recognition.
• To enhance the performance of the YOLOv11n model through targeted optimizations.
• To enable real-time pig behavior detection for smart livestock management.

Main Methods:

• Proposed three optimizations for the lightweight YOLOv11n model: SCSA-CBAM for feature discrimination, WFU for cross-scale integration, and WTConv for reduced computational overhead.
• Employed a two-stage transfer learning approach with data augmentation, initialized with COCO pre-trained weights.
• Evaluated the model on a self-built six-category pig behavior dataset (2480 images).

Main Results:

• The optimized YOLOv11n-SCSA-WFU-WT model achieved high performance metrics (mAP@0.5: 0.974, mAP@0.5:0.95: 0.785).
• The model demonstrates a favorable accuracy-efficiency trade-off with 3.40 M parameters, 7.8 GFLOPs, and 72.28 FPS.
• Ablation studies confirmed the effectiveness of each proposed module, showing substantial improvements over the baseline.

Conclusions:

• The proposed optimized YOLOv11n model effectively facilitates real-time pig behavior detection.
• The method offers a lightweight yet accurate solution suitable for resource-limited smart livestock management systems.
• The advancements contribute to improved early disease detection and animal welfare monitoring in pigs.