Automatic Perception of Typical Abnormal Situations in Cage-Reared Ducks Using Computer Vision.

Area of Science:

• Animal Science
• Computer Vision
• Agricultural Technology

Background:

• Overturning and death are significant welfare issues in cage-reared ducks.
• Early detection of these abnormalities is crucial for timely intervention and improved animal welfare.
• Existing detection methods may lack accuracy and robustness, especially under varying conditions.

Purpose of the Study:

• To develop an accurate and robust deep learning model for detecting overturning and death in cage-reared ducks.
• To enhance the YOLOv8 object detection algorithm by incorporating attention mechanisms and an improved loss function.
• To evaluate the model's performance, generalization ability, and efficiency in real-world conditions.

Main Methods:

• Modified YOLOv8 incorporating GAM attention mechanisms in the feature fusion neck.
• Implementation of the Wise-IoU loss function to balance data samples and reduce geometric parameter penalties.
• Pose estimation using HRNet-48 focusing on six key body points for refined posture analysis.
• Testing with adjusted image brightness (0.85, 1.25) and comparison with mainstream object-detection algorithms.

Main Results:

• The proposed model achieved a mean Average Precision (mAP) of 0.924, outperforming the original YOLOv8 by 1.65%.
• The model demonstrated excellent generalization ability and robustness against lighting variations.
• The pose-estimation model achieved an Object Key point Similarity (OKS) of 0.921 with a processing time of 0.528s per frame.

Conclusions:

• The enhanced YOLOv8 model effectively detects overturning and death abnormalities in cage-reared ducks with high accuracy.
• The integration of GAM attention and Wise-IoU significantly improves detection performance and robustness.
• The developed pose-estimation model accurately identifies abnormal postures, offering a valuable tool for duck welfare monitoring.