TL;DR
This paper introduces the first non-invasive, automated computer vision system designed to determine the gender of day-old chicks using facial images. By adapting human facial gender classification techniques, the proposed end-to-end framework achieves an accuracy of 81.89%, offering a scalable, animal-welfare-friendly alternative to traditional, invasive industrial methods.
What Problem Does This Paper Address?
In the poultry industry, identifying the sex of day-old chicks is a critical operational step because male and female chickens serve entirely different commercial roles (e.g., egg production vs. meat yield).
Current industrial methods suffer from significant trade-offs. Feather and color sexing are fast and non-invasive but only work on specific genetic breeds. Vent sexing—which involves manually examining the chick's cloaca—is universally applicable and highly accurate (~97%), but it is highly invasive, risks injuring the bird, and requires specialized human experts who undergo up to 18 months of intensive training.
To bridge this gap, the authors explore whether deep learning can detect subtle, non-invasive visual cues from a chick's face to automate gender classification, minimizing animal handling and eliminating the need for specialized human expertise.
Methods
The authors develop a multi-stage computer vision pipeline that mirrors human facial recognition workflows:
- Chick Face Localization: A pre-trained YOLOv5 object detection model is fine-tuned to locate and draw bounding boxes around the chick's face.
- Keypoint Detection: A High-Resolution Network (HRNet) is customized to detect 7 key facial landmarks: the eyes, upper nose, middle nose, and three points across the beak.
- Face Alignment: Using geometric calculations derived from the eye coordinates, the system automatically rotates and scales the image to ensure the chick's face is in a uniform, frontal position.
- Data Preparation (Cropping Strategies): The authors experiment with two input types: Cropped Full Face (the entire aligned face) and Cropped Middle Face (a tighter crop isolating just the eyes and beak using image binarization thresholding).
- Feature Extraction & Classification: Six standard deep learning backbones (AlexNet, EfficientNet-B0, Inception-V3, ResNet-50, ResNet-101, and VGG-16) are evaluated alongside a Multi-Layer Perceptron (MLP) classifier to output the final gender prediction.
Datasets
The study introduces a newly collected, specialized dataset supported by Cobb-Vantress poultry facilities:
- Total Images: 5,727 high-quality facial images.
- Chicks Sampled: 353 day-old chicks (184 female and 169 male).
- Annotations: Meticulously annotated by two poultry experts and two computer vision experts to establish ground truth for face bounding boxes, facial keypoints, and gender.
Key Insights and Ideas
- The Beak and Comb Hold the Secret: Using Grad-CAM++ visualizations, the authors found that the models consistently focused on the region around the chick's beak (specifically the middle nose keypoint) and the comb area. This confirms that deep learning models can pick up on subtle, biologically dimorphic facial features that are virtually imperceptible to human sorters.
- More Context Outperforms Tighter Crops: Counterintuitively, zooming in too close hurts performance. The Cropped Full Face strategy consistently outperformed the Cropped Middle Face strategy. Tight cropping likely discards peripheral features, such as head shape or comb structure, that contribute to gender differentiation.
- Architecture Complexity Matters: Residual networks (ResNet-50 and ResNet-101) significantly outperformed simpler architectures like VGG-16, demonstrating that multi-scale feature extraction is necessary to map these highly nuanced facial differences.
Key Results
- Top Classification Performance: ResNet-50 achieved the highest overall accuracy of 81.89% on the Cropped Full Face dataset, yielding 81.31% precision, 81.85% recall, and an AUC of 89.51%.
- Cropping Comparison: On the tighter Cropped Middle Face dataset, ResNet-50 still led but with a lower top accuracy of 80.46%.
- Localization Robustness: The initial YOLOv5 face localization step was highly effective, achieving 99% accuracy in bounding box predictions on unseen images.
Limitations and Future Work
- Dataset Scale and Diversity: While promising as a proof-of-concept, a dataset of 353 chicks is relatively small. Expanding the dataset to include more diverse chicken breeds and larger sample sizes is required to improve model robustness.
- Environmental Sensitivity: The system's accuracy can be influenced by real-world variations in lighting, background clutter, and motion blur from un-still chicks.
- Industrial Throughput Integration: For commercial viability, this software pipeline must be integrated into high-speed automated hatchery machinery—such as vaccination machines that already hold chicks by the head—requiring real-time processing capabilities under rapid throughput conditions.
Summary updated