Pattern-Based Felidae Subspecies Classification Through Transfer and Ensemble Learning Methods

This paper proposes a deep learning-based approach using Convolutional Neural Networks (CNNs) to automatically classify Felidae species. This study focuses on five major species: Leopard, Cheetah, Jaguar, Tiger, and Lynx. A dataset of high-resolution images representing these species were used, and a CNN model was implemented using Python. The proposed model is designed with a deep convolutional architecture using Transfer Learning (TL) from pre-trained networks, ensuring efficient feature extraction and classification. Specifically, six state-of-the-art CNN architectures were used: EfficientNet-B0, ConvNet-Tiny, ViT-Small (Vision Transformer Small), GoogleNet, denseNet, and Xception. To assess performance, Ensemble Learning (EL) techniques that combine models have been applied for more accurate and robust predictions. These models are trained and evaluated to determine the most effective architecture for the classification of Felidae species. Extensive data augmentation techniques are used to improve model generalization. The model achieves high classification accuracy, demonstrating the effectiveness of deep learning in taxonomic classification tasks. Visualizations of the model’s performance, including accuracy curves, confusion matrices, and feature activation maps, are provided to illustrate its effectiveness. Compared to other EL techniques, the Voting classifier exhibits strong overall performance, achieving 97% of accuracy, 97% of precision, 97% of F1-scores and also 97% recall in multi-class animal classification over 20 epochs. These results underscore the Voting model’s robustness and reliability across diverse felidae species. This study contributes to automated wildlife monitoring and conservation efforts by improving the precision and efficiency of Felidae species classification.