An automatic classifier for monitoring applied behaviors of cage-free laying hens with deep learning

Poultry behavior is an important indicator of their welfare, health, and production performance. The welfare of layers and broilers such as walking ability, breast blisters, hock burn, and heart failures are measurable through behavior monitoring. In the previous research, most of laying hen studies focused on basic behaviors such as drinking, feeding, and walking of broilers. However, with the transition to the cage-free houses, more natural behaviors need to be monitored for welfare assessment. In this study, a six-behavioral classifier (i.e., feeding, drinking, walking, perch, dust bathing, and nesting) was developed based on multiple CNN models (e.g., efficientNetV2 and YOLOv5-cls). The classifier is one of the first model included perching, dust bathing, and nesting behaviors, which are special characters that reflect basic welfare of cage-free birds. Furthermore, a cage-free birds’ dataset containing 12,000 pictures was collected and annotated in a lifespan scale (e.g., from 1 week to 50 weeks of old), from which 9,600 images were used as training dataset and the rest were used for validation. The best performance model YOLOv5-cls-m achieved an average accuracy of 95.3%, which is 5.01% higher than that of efficientNetV2-l. Drinking behavior of chicks was monitored with the highest accuracy (97.8%) while nesting behavior had a detection precision of 92.5%. In terms of chickens’ age, the classifier has a better accuracy for smaller chicks (< 10 days) than larger chickens older than 10 days (96.4% vs 94.3%). The results show that the classifier is a useful tool to segregate cage-free bird behaviors in various life periods and environments.