Sheep face identification using multi-model deep learning evaluation and ensemble learning

The individual identification of sheep is critically important for herd management, health monitoring, and record accuracy. The limitations of traditional identification methods-such as ear tags, markings, and Radio Frequency Identification (RFID)-have accelerated the adoption of image-based biometric approaches. In recent years, deep learning-based facial identification systems have emerged as a prominent alternative in modern livestock management due to their contactless, fast, and user-friendly nature. Despite these advances, studies focusing specifically on individual sheep identification remain limited, and systematic comparisons of modern architectures and ensemble strategies are still scarce. To address this gap, this study reconstructs an existing dataset for individual sheep identification and systematically evaluates state-of-the-art deep learning architectures and ensemble approaches.

A dataset originally developed for breed classification was reconstructed for individual sheep face identification. After data cleaning, the dataset comprised 1680 facial images from 172 unique sheep. To increase data diversity, comprehensive augmentation strategies were applied, expanding the dataset to 14,810 images. Sixteen deep learning architectures-including EfficientNet (B0, B1, B2, B3, B5), DenseNet (121, 169, 201), MobileNet (V2, V3-Small, V3-Large), ResNet50, InceptionV3, NASNetMobile, Xception, and VGG16-were evaluated under both baseline and fine-tuned configurations. Based on the fine-tuned results, the three highest-performing models were selected to develop three ensemble approaches, namely Soft Ensemble, Weighted Ensemble, and Stacked Ensemble. In addition, a user-friendly graphical user interface (GUI) was implemented to support practical sheep identification from user-uploaded images.

Under baseline configurations, EfficientNetB1, EfficientNetB2, and DenseNet201 achieved identification accuracies of 85.0 %, 84.5 %, and 87.7 %, respectively. After fine-tuning, substantial performance improvements were observed, with accuracies increasing to 91.3 % for EfficientNetB1, 91.0 % for EfficientNetB2, and 90.7 % for DenseNet201. Using these fine-tuned models, ensemble strategies further enhanced identification performance. The Soft Ensemble achieved an accuracy of 92.4 %, the Weighted Ensemble reached 92.6 %, and the Stacked Ensemble achieved the highest accuracy of 93.7 %. The developed GUI enabled fast and intuitive verification of user-uploaded face images, supporting real-time on-farm identification workflows.

This study provides a comprehensive evaluation of modern deep learning architectures for individual sheep face identification and demonstrates that fine-tuning significantly improves model performance. Furthermore, the proposed Stacked Ensemble approach achieves the highest and most balanced identification performance among all evaluated methods. The results confirm that facial biometrics, supported by ensemble-driven deep learning and a practical GUI, offer a reliable and scalable solution for livestock identification. Future work will focus on expanding the dataset, improving robustness across diverse farm environments, and optimizing the framework for real-time deployment on resource-constrained edge devices.