EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Entropy-Weighted Geometry for Robust Vision Embeddings

P. Karthiksankar, A. Pandiarani

Boletim da Sociedade Paranaense de Matemática · 03.07.2026 · 44 (7) · 1-23

DOI: 10.5269/bspm.82953 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Boletim da Sociedade Paranaense de Matemática, online erschienen am 03.07.2026. Verfasst von P. Karthiksankar und 1 weiteren Autor:innen. Die Arbeit nennt 22 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This paper introduces an entropy-weighted geometric framework for analyzing the stability of deep vision embeddings under stochastic perturbations. The proposed approach replaces classical Euclidean geometry with a nonlinear entropy-based distance that modulates coordinate contributions according to their magnitude. We show theoretically that entropy-weighted distortions grow more slowly than Euclidean distortions under Gaussian perturbations, leading to improved robustness of embedding representations. An algorithmic framework for computing entropy-weighted distances is developed and applied to feature embeddings extracted from pretrained convolutional neural networks. Extensive experiments evaluate distortion growth, margin stability, and global geometric structure under controlled noise perturbations. The results demonstrate that entropy-weighted geometry consistently produces smaller distortions and preserves structural relationships between embeddings more effectively than Euclidean distance. Heatmap visualizations further reveal a clearer global organization of embedding spaces under the proposed metric. These findings suggest that entropy-weighted geometry provides a principled and computationally efficient framework for studying robustness properties of high-dimensional representation spaces in modern vision systems.

Zitieren

APA

Karthiksankar, P., & Pandiarani, A. (2026). Entropy-Weighted Geometry for Robust Vision Embeddings. Boletim da Sociedade Paranaense de Matemática, 44(7), 1-23. https://doi.org/10.5269/bspm.82953

Vancouver

Karthiksankar P, Pandiarani A. Entropy-Weighted Geometry for Robust Vision Embeddings. Boletim da Sociedade Paranaense de Matemática. 2026;44(7):1-23. doi:10.5269/bspm.82953

BibTeX
@article{Karthiksankar2026,
  author = {Karthiksankar, P. and Pandiarani, A.},
  title = {Entropy-Weighted Geometry for Robust Vision Embeddings},
  journal = {Boletim da Sociedade Paranaense de Matemática},
  year = {2026},
  volume = {44},
  number = {7},
  pages = {1--23},
  doi = {10.5269/bspm.82953}
}
Kostenlos starten →