The primary objective of this project was to develop a robust and accurate machine learning model for the classification of various animal species from images.
Furthermore, using the power of transfer learning, this research sought to create a sophisticated multi-class classification system capable of distinguishing and categorizing different animal species based on their visual features.
A key component of this thesis was the evaluation and implementation of transfer learning techniques. Transfer learning involved using pre-trained models as a starting point, which significantly reduced the time and computational resources required for training, while potentially improving the model's accuracy and generalization capabilities.
A comparative analysis of various pre-trained models, such as those based on deep learning architectures like Convolutional Neural Networks (CNNs), was conducted.
This analysis focused on different aspects of the models, including architecture choice, fine-tuning strategies, and the impact of various layers and weights on the classification task.
The models were assessed based on multiple performance metrics, including accuracy, precision, recall, and confusion matrix. By exploring these dimensions, the goal was to establish a model that not only delivered high classification accuracy but also balanced the trade-offs between performance and resource utilization.
• Custom CNN model.
• Pre-Trained Model:
1. VGG19
2. ResNet50
3. MobileNetV2
4. InceptionV3
• Cross Validation
• Evaluation Metrics
The results indicate that transfer learning, especially with models like MobileNetV2 and InceptionV3, offers superior performance in terms of accuracy, generalization, and computational efficiency compared to a traditional CNN model.
MobileNetV2 emerged as the best-performing model, achieving near-perfect validation and testing accuracies (99.9% and 99%, respectively), followed closely by InceptionV3, which also showed robust performance across all folds.
The Base CNN model demonstrated good training accuracy but suffered from overfitting, as reflected by its drop in testing accuracy. VGG19 offered solid performance but was outperformed by the more computationally efficient MobileNetV2 and InceptionV3.
The ResNet50 model performed poorly, with inconsistent results across different folds and low testing accuracy, suggesting it was not well-suited for this specific task without further fine-tuning.
Overall, this analysis highlights the effectiveness of transfer learning in multi-class animal classification tasks, particularly when using lightweight models like MobileNetV2, which balance high accuracy with computational efficiency.