A Comparative Evaluation Of Deep Learning Architectures For Butterfly Species Classification In Natural Scene Images

Automated ecological species identification is a subject of intense study in computer vision research, which plays an important role in biodiversity monitoring, research, and conservation planning. The research work presented in this article introduces a large custom butterfly dataset consisting of 23537 images belonging to 102 different species, collected through extensive field photography and supplemented with selected images from reliable online sources. Furthermore, the dataset is preprocessed with annotations to generate a Region of Interest (RoI). A total of 54 state-of-the-art deep learning architectures, including classical Convolution Neural Networks (CNN), Modern CNN, Lightweight CNN models, attention-based models, and Vision Transformers, were evaluated on both segmented and non-segmented images of the dataset. A comparative analysis of all the deep learning models in terms of accuracy, precision, recall and F1-Score has been done. The non-uniform class distribution in the dataset motivates the use of weighted evaluation metrics along with overall accuracy, ensuring an impartial evaluation across all categories. Results indicate that modern CNN architectures consistently outperform other models on both segmented and non-segmented images in the dataset. In particular, ConvNeXt-Tiny achieved the best performance of a weighted F1-Score of 90.88% on segmented images and 84.8% on non-segmented images. This proves the effectiveness of object-focused preprocessing and demonstrates the superiority of modern CNN architectures for fine-grained butterfly species classification.