Abstract
Automated recognition of animal species from images remains a central challenge in computer vision with applications in ecological monitoring, veterinary diagnostics, and livestock management. Conventional convolutional architectures exhibit limited robustness when confronted with heterogeneous data and visually similar classes, resulting in reduced inter-class separability. To address this limitation, three representative deep learning models were examined: a custom convolutional network, a deep residual network, ResNet50, and a transformer-based ViT-B16. All models were trained under identical conditions with standardized normalization, extensive augmentation, and uniform optimization. Comparative evaluation established a clear stratification of performance. The custom network achieved only moderate accuracy with pronounced misclassification between similar categories. ResNet50 demonstrated substantial improvements and more stable generalization, though residual ambiguities persisted. ViT-B16 consistently outperformed both convolutional models, achieving class-wise F1-scores above 0.90 and true positive rates exceeding 0.95 for most categories. The findings provide rigorous evidence of the superiority of transformer-based architectures for fine-grained image recognition.