10th International Congress on Information and Communication Technology

Authors - Kanda Tshinu Patrick, Chunling Tu, Owolawi Pius Adewale, Antonie Smith
Abstract - This paper presents a comparative study for animal classification using Convolutional Neural Networks (CNNs), specifically based on the VGG16, AlexNet, and a custom CNN architecture. Traditional methods for classifying animals rely heavily on human visual ability, which is time-consuming, labor-intensive, and prone to inconsistencies that can compromise the accuracy of species identification. In contrast, CNN-based techniques leverages automated features extraction. Primarily focusing on body and facial features, to deliver more consistent and reliable classification outcomes. The fields of computer vision and image processing have gained significant attention due to their effectiveness in addressing classification challenges across domains such as agriculture, wildlife conservation, and biodiversity research. The development of CNNs has become a critical tool in these areas, enabling automated systems to monitor livestock, track animal behaviour, and detect wildlife, enhancing both animal management and conservation efforts. In this study, three CNN architectures—VGG16, AlexNet, and a custom CNN—were trained on a dataset comprising 5,400 images of animals across 90 classes. The results show that VGG16 achieved the highest classification accuracy of 91.51%, followed by the custom CNN with 84.57%, and AlexNet with 79.02%. These findings demonstrate the potential of deep learning for accurate species identification. Furthermore, the use of pre-trained models like VGG16 can enhance classification performance, while custom models can provide competitive results with fewer computational resources. This study highlights the effectiveness of CNNs in animal classification and underscores their potential for discovering new observations, including the identification of previously undocumented species within the same class.