Animal Image Classification Using DenseNet-161 Transfer Learning: A Comprehensive Deep Learning Approach for Multi-Species Recognition

This paper details a deep learning method for automatically classifying animal species. It uses the DenseNet-161 architecture along with transfer learning techniques. The study tackles the important challenge of classifying images into multiple fine-grained categories across different taxonomic groups. The model was trained on a carefully selected dataset of 5,400 images with a balanced class distribution (60 images per class) of 90 different animal species. After six training epochs with early stopping, the model achieved validation accuracy of 89.11% and showed strong generalization with 85.00% test accuracy on 360 held-out images. Utilizing a comprehensive data augmentation pipeline that included random cropping, horizontal flipping, Gaussian blur, color jitter, and rotation transformations, the implementation made use of the PyTorch deep learning framework. Training took about 49 minutes on a CUDA-enabled GPU. The result suggests transfer learning is suitable for multi-class image classification by freezing convolutional backbone layers where DenseNet-161 appears to be efficient with parameters. They highlight the benefits of dense connectivity and show that transfer learning works well for specific computer vision tasks in wildlife conservation, ecological monitoring, biodiversity assessment, and automated species identification systems. Perclass analysis showed both the advantages and disadvantages of the transfer learning approach with limited training data, revealing perfect classification (100% F1-score) on 40% of species while highlighting particular difficulties with morphologically similar small mammals.