Enhancing Out-of-Distribution Detection Under Covariate Shifts: A Full-Spectrum Contrastive Denoising Framework DOI Open Access
Dengye Pan, Bin Sheng, Xiaoqiang Li

и другие.

Electronics, Год журнала: 2025, Номер 14(9), С. 1881 - 1881

Опубликована: Май 6, 2025

Out-of-distribution (OOD) detection is crucial for identifying samples that deviate from the training distribution, thereby enhancing reliability of deep neural network models. However, existing OOD methods primarily address semantic shifts, where an image’s inherent semantics have changed, and often overlook covariate which are prevalent in real-world scenarios. For instance, variations image contrast, lighting, or viewpoints can alter input features while keeping content intact. To this, we propose Full-Spectrum Contrastive Denoising (FSCD) framework, improves under shifts. FSCD first establishes a robust boundary then refines feature representations through fine-tuning. Specifically, employs dual-level perturbation augmentation module to simulate shifts contrastive denoising effectively distinguish in-distribution samples. Extensive experiments on three benchmarks demonstrate achieves state-of-the-art performance, with AUROC improvements up 0.51% DIGITS, 0.55% OBJECTS, 2.09% COVID compared previous best method also maintaining highest classification accuracy covariate-shifted

Язык: Английский

Enhancing Out-of-Distribution Detection Under Covariate Shifts: A Full-Spectrum Contrastive Denoising Framework DOI Open Access
Dengye Pan, Bin Sheng, Xiaoqiang Li

и другие.

Electronics, Год журнала: 2025, Номер 14(9), С. 1881 - 1881

Опубликована: Май 6, 2025

Out-of-distribution (OOD) detection is crucial for identifying samples that deviate from the training distribution, thereby enhancing reliability of deep neural network models. However, existing OOD methods primarily address semantic shifts, where an image’s inherent semantics have changed, and often overlook covariate which are prevalent in real-world scenarios. For instance, variations image contrast, lighting, or viewpoints can alter input features while keeping content intact. To this, we propose Full-Spectrum Contrastive Denoising (FSCD) framework, improves under shifts. FSCD first establishes a robust boundary then refines feature representations through fine-tuning. Specifically, employs dual-level perturbation augmentation module to simulate shifts contrastive denoising effectively distinguish in-distribution samples. Extensive experiments on three benchmarks demonstrate achieves state-of-the-art performance, with AUROC improvements up 0.51% DIGITS, 0.55% OBJECTS, 2.09% COVID compared previous best method also maintaining highest classification accuracy covariate-shifted

Язык: Английский

Процитировано

0