Mask-Aware DINOv2 Foundation Features for Label-Efficient Environmental Microorganism Classification: A Proof-of-Concept Study on Environmental Microorganism Image Dataset Sixth Version
DOI:
https://doi.org/10.21928/uhdjst.v10n2y2026.pp226-23Keywords:
Environmental Microorganisms, DINOv2, Vision Transformers, Support Vector Machine, Environmental Microorganism Image Dataset Sixth Version-6Abstract
Environmental microorganism classification is difficult because curated microscopy datasets are small, visual appearance varies within classes, and expert annotation is costly. We examined mask-guided frozen DINOv2 representations on environmental microorganism image dataset sixth version (EMDS-6) using 840 images from 21 balanced classes. Twenty repeated partitions were stratified by class and grouped by exact image hash so that identical content could not cross training, validation, and test sets. The global DINOv2 class-token baseline and full hybrid were evaluated with 1–24 labeled training images per class; DINOv2-only, handcrafted-only, partial-fusion, full-fusion, and leave-one-block-out ablations were evaluated at 24 shots. Ground-truth masks were used for every mask-guided feature, so this remains an oracle spatial-guidance study rather than an end-to-end system. At 24 shots, the full hybrid linear support vector machine achieved 95.71% ± 1.23% accuracy and 95.67% ± 1.26% macro-F1, compared with 92.93% ± 1.82% macro-F1 for global DINOv2. The paired gain was 2.74 percentage points (95% bootstrap confidence interval, 1.86–3.61; exact signed-rank P = 0.000019; Holm-adjusted P across six label budgets = 0.000057). However, global-plus-crop features reached 95.33% ± 1.40% macro-F1 and did not differ significantly from the full hybrid after correction. Handcrafted features alone reached 64.00% ± 3.44%, and removing them from the hybrid did not reduce performance. The results support mask-guided frozen features within EMDS-6, but not deployment, segmentation robustness, supervised-model superiority, or external generalizability.
References
[1] P. Zhao, C. Li, M. M. Rahaman, H. Xu, P. Ma, H. Yang, H. Sun, T. Jiang, N. Xu and M. Grzegorzek. “EMDS-6: Environmental microorganism image dataset sixth version for image denoising, segmentation, feature extraction, classification, and detection method evaluation”. Frontiers in Microbiology, vol. 13, p. 829027, 2022.
[2] P. Zhao, C. Li, M. M. Rahaman, H. Xu, P. Ma, H. Yang, H. Sun, T. Jiang, N. Xu and M. Grzegorzek. “EMDS-6.” Figshare, United Kingdom, 2021.
[3] P. Zhao, C. Li, M. M. Rahaman, H. Xu, H. Yang, H. Sun, T. Jiang and M. Grzegorzek. “A comparative study of deep learning classification methods on a small environmental microorganism image dataset (EMDS-6): From convolutional neural networks to visual transformers”. Frontiers in Microbiology, vol. 13, p. 792166, 2022.
[4] J. Yuan, P. Zhao, N. Xu, C. Li and T. Jiang. “A Few-Shot Learning Based Environmental Microorganism Image Classification Method”. In: Proceedings 2024 2nd International Conference on Machine Vision, Image Processing and Imaging Technology (MVIPIT). pp. 199-203, 2024.
[5] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit and N. Houlsby. “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”. In: Proceedings International Conference Learning Representations (ICLR). arXiv, New York, 2021.
[6] M. Oquab, T. Darcet, T. Moutakanni, H. V. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes.,… & P. Bojanowski. “DINOv2: Learning robust visual features without supervision”. Transactions on Machine Learning Research, 2024. Available: https://openreview.net/forum?id=a68SUt6zFt
[7] M. K. Hu. “Visual pattern recognition by moment invariants”. Transactions on Information Theory, vol. 8, no. 2, pp. 179-187, 1962.
[8] T. Ojala, M. Pietikäinen and T. Mäenpää. “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 971-987, 2002.
[9] R. M. Haralick, K. Shanmugam and I. Dinstein. “Textural features for image classification”. IEEE Transactions on Systems, Man, and Cybernetics, vol. SMC-3, no. 6, pp. 610-621, 1973.
[10] C. M. Liang, C. C. Lai, S. H. Wang and Y. H. Lin. “Environmental microorganism classification using optimized deep learning model”. Environmental Science and Pollution Research, vol. 28, no. 24, pp. 31920-31932, 2021.
[11] M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski and A. Joulin. “Emerging Properties in Self-Supervised Vision Transformers”. In: Proceedings IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650-9660, 2021.
[12] K. He, X. Chen, S. Xie, Y. Li, P. Dollár and R. Girshick. “Masked autoencoders are scalable vision learners. In: Proceedings IEEE/ CVF Conference Computer Vision and Pattern Recognition (CVPR)”. pp. 16000-16009, 2022.
[13] M. Nielsen, L. Wenderoth, T. Sentker and R. Werner. “Self-supervision for medical image classification: State-of-the-art performance with ~100 labeled training samples per class”. Bioengineering (Basel), vol. 10, no. 8, p. 895, 2023.
[14] D. Kang, P. Koniusz, M. Cho and N. Murray. “Distilling Self- Supervised Vision Transformers for Weakly-Supervised Few-Shot Classification and Segmentation”. In: Proceedings IEEE/CVF Conference Computer Vision and Pattern Recognition (CVPR). pp. 19627-19638, 2023.
[15] F. Kulwa, C. Li, J. Zhang, K. Shirahama, S. Kosov, X. Zhao, T. Jiang and M. Grzegorzek. “A new pairwise deep learning feature for environmental microorganism image analysis”. Environmental Science and Pollution Research, vol. 29, no. 34, pp. 51909-51926, 2022.
[16] L. Nanni, E. De Luca, M. L. Facin and G. Maguolo. “Deep learning and handcrafted features for virus image classification”. Journal of Imaging, vol. 6, no. 12, p. 143, 2020.
[17] H. Zheng, R. Wang, Z. Yu, N. Wang, Z. Gu and B. Zheng. “Automatic plankton image classification combining multiple view features via multiple kernel learning”. BMC Bioinformatics, vol. 18, no. Suppl 16, p. 570, 2017.
[18] T. Panaïotis, E. Amblard, G. Boniface-Chang, G. Dulac-Arnold, B. Woodward and J. O. Irisson. “Benchmark of plankton images classification: Emphasizing features extraction over classifier complexity”. Earth System Science Data, vol. 18, pp. 945-967, 2026.
[19] C. Cortes and V. Vapnik. “Support-vector networks”. Machine Learning, vol. 20, pp. 273-297, 1995.
[20] C. Guo, G. Pleiss, Y. Sun and K. Q. Weinberger. “On Calibration of Modern Neural Networks”. In: Proceedings 34th International Conference Machine Learning (ICML). Proceedings of Machine Learning Research, vol. 70, pp. 1321-1330, 2017.
[21] Y. Romano, M. Sesia and E. J. Candès. “Classification with Valid and Adaptive Coverage”. In: Proceedings Advances in Neural Information Processing Systems 33, 2020.
[22] S. Holm. “A simple sequentially rejective multiple test procedure”. Scandinavian Journal of Statistics, vol. 6, no. 2, pp. 65-70, 1979.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Abdulsalam Abdulla Ghadir, Chewan Jalal Mohammed, Mzhda Hiwa Hama
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
