Deep learning-based facial recognition for identity and sex classification in chicks

To address the digitalization of poultry farming and demand for precise chick management, this study developed deep learning models for facial biometric recognition in Xianghuang chickens, an indigenous yellow-feathered breed from Hengyang, Hunan. A unified dataset of 6,024 images (frontal and profile views) was compiled to evaluate five architectures-ResNet-50, ResNet-101, MobileNetV2, EfficientNet-B0, and EfficientNet-B0+CBAM-on two tasks: individual identification and sex classification. For individual identification, all models performed excellently. Under frontal views, Top-1 accuracy exceeded 98% and Top-5 surpassed 99%, with no significant inter-model differences (p> 0.05). Profile views were slightly more challenging, but Top-1 accuracy remained above 94%. Grad-CAM visualization revealed models focused on key facial regions (nasal, periocular, beak), confirming facial features' utility for identification. Sex classification using facial features alone was unsatisfactory. The highest accuracy was 71.01% (ResNet-50, profile view), with all models showing overfitting (generalization gap > 2%) and stagnant validation accuracy. This stems from subtle sexual dimorphism in slow-growing chicks and imaging variability. This study constructed the first labeled facial dataset for Xianghuang chickens, filling a gap in poultry facial recognition research. While demonstrating deep learning's potential for non-contact chick identification under controlled conditions, it highlights biological limitations in early sex determination via facial morphology alone.