Butterflies and benchmarks: a systematic evaluation of deep learning backbones to identify the best model for image-based butterflies species recognition
Anwar H (2026)
Publication Type: Journal article
Publication year: 2026
Journal
Book Volume: 29
Article Number: 13
Journal Issue: 1
DOI: 10.1007/s10044-025-01587-7
Abstract
Recent advancements in artificial intelligence (AI), particularly deep learning, offer promising tools that can support the biodiversity conservation efforts in the face of challenges like climate change and carbon emissions. The paper extends this theme by focusing on butterflies often regarded as nature’s living canvases, and aims to achieve their image-based species recognition using advanced AI algorithms, specifically convolutional neural networks (CNNs) and vision transformers (ViTs). However, the performance of these algorithms is severely affected by the variations found in butterflies images due to differences in scale, pose, background clutter, body deformations, and illumination. To this end, most prior approaches rely on existing architectures, extend them or reuse their blocks. Except for a few that evaluated selected architectures, no study has comprehensively benchmarked available models on public butterfly datasets to identify the one best suited for image-based butterfly species recognition in terms of both accuracy and efficiency. To fill this gap, the paper evaluates 65 CNNs and 65 ViTs on two publicly available butterfly image datasets covering over 100 species. The images are first encoded into feature vectors using the versions of these models pre-trained on large-scale generic image dataset. For each model, the resulting feature vectors are then used to train and evaluate a linear Support Vector Machine (SVM) via 5-fold cross-validation. The use of a simple classifier ensures that performance reflects the quality of the extracted features, not the classifier itself, while the SVM training time reflects the computational efficiency of the encoders. The overall benchmarking results indicate that ViTs perform better than CNNs both with respect to accuracy and efficiency. The code and links to both the image datasets are available at https://github.com/hafeez-anwar/Butterflies-Benchmark.
Involved external institutions
Pakistan (PK)How to cite
APA:
Anwar, H. (2026). Butterflies and benchmarks: a systematic evaluation of deep learning backbones to identify the best model for image-based butterflies species recognition. Pattern Analysis and Applications, 29(1). https://doi.org/10.1007/s10044-025-01587-7
MLA:
Anwar, Hafeez. "Butterflies and benchmarks: a systematic evaluation of deep learning backbones to identify the best model for image-based butterflies species recognition." Pattern Analysis and Applications 29.1 (2026).
BibTeX: Download