Tracking perching behavior of cage-free laying hens with deep learning technologies

Area of Science:

• Animal Science
• Computer Science
• Agricultural Engineering

Background:

• Cage-free (CF) housing offers laying hens benefits like improved bone health and reduced stress.
• Accurate detection of perching behavior is crucial for monitoring hen welfare and development in CF systems.
• Manual observation for perching behavior is labor-intensive and can be inaccurate.

Purpose of the Study:

• To develop and test a deep learning model for detecting perching behavior in laying hens.
• To evaluate the optimal model's performance across different ages of hens in CF housing.

Main Methods:

• Developed and compared deep learning models (YOLOv8s-PB, YOLOv8x-PB, YOLOv7-PB, YOLOv7x-PB) for perching behavior detection.
• Trained models using 3,000 images from 4 CF rooms with perches up to 1.8m high.
• Evaluated model accuracy, precision, recall, and mAP@0.50 using 1-way ANOVA.

Main Results:

• The YOLOv8x-PB model demonstrated superior performance with 94.80% precision, 95.10% recall, and 97.60% mAP@0.50.
• All tested models achieved over 94% detection precision.
• Detection precision varied by age phase, with the peaking phase showing the highest (97.40%) and starter phase the lowest (88.80%).
• Overlapping birds and occlusion slightly reduced detection performance.

Conclusions:

• The YOLOv8x-PB model is highly effective for automated detection of laying hen perching behavior.
• This deep learning tool can assist CF producers in monitoring hen welfare and developmental stages.
• Accurate monitoring of perching behavior supports better management practices in cage-free systems.