EUVIMEDEUVIMED entdecken
Crossref

Local Background Features Matter in Out-of-Distribution Detection

Jinlun Ye, Zhuohao Sun, Yiqiao Qiu, Qiu Li, Zhijun Tan, Ruixuan Wang

Neural Computation · 05.11.2026 · 1-39

DOI: 10.1162/neco.a.1601 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Neural Computation, online erschienen am 05.11.2026. Verfasst von Jinlun Ye und 5 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Out-of-distribution (OOD) detection is crucial when deploying deep neural networks in the real world to ensure the reliability and safety of their applications. One main challenge in OOD detection is that neural network models often produce overconfident predictions on OOD data. Existing methods address this issue by using auxiliary OOD data sets, generating fake OOD images, or synthesizing pseudo-OOD features, but these methods often require additional data sources, generation pipelines, or extra training components. In this study, we propose a local background feature fine-tuning framework for OOD detection. Instead of constructing explicit pseudo-OOD images or synthesizing artificial outlier features, our method directly extracts local background features from ID images and uses them as feature-level regularization targets. Specifically, local features with low ground-truth class probability are selected as background-like features, and a margin-based L2-norm loss is applied to suppress their activation magnitude during fine-tuning. Together with the cross-entropy loss that preserves foreground discriminability, this design reduces background-induced overconfidence without requiring auxiliary OOD data, generated images, segmentation models, or additional OOD-specific branches. Extensive experiments on standard OOD detection benchmarks demonstrate the effectiveness of our method and its compatibility with existing post hoc scoring functions, achieving competitive or state-of-the-art performance.

Zitieren

APA

Ye, J., Sun, Z., Qiu, Y., Li, Q., Tan, Z., & Wang, R. (2026). Local Background Features Matter in Out-of-Distribution Detection. Neural Computation, 1-39. https://doi.org/10.1162/neco.a.1601

Vancouver

Ye J, Sun Z, Qiu Y, Li Q, Tan Z, Wang R. Local Background Features Matter in Out-of-Distribution Detection. Neural Computation. 2026:1-39. doi:10.1162/neco.a.1601

BibTeX
@article{Ye2026,
  author = {Ye, Jinlun and Sun, Zhuohao and Qiu, Yiqiao and Li, Qiu and Tan, Zhijun and Wang, Ruixuan},
  title = {Local Background Features Matter in Out-of-Distribution Detection},
  journal = {Neural Computation},
  year = {2026},
  pages = {1--39},
  doi = {10.1162/neco.a.1601}
}
Kostenlos starten →