غير محددة في المصدر 2026 بحاجة إلى مراجعة

UnU-Net: Distance-Aware Uncertainty Quantification for Semantic Segmentation

زيارة المصدر الأصلي

الملخص

Automated mechanical inspections increasingly rely on deep learning models for semantic segmentations of faults. However, standard deterministic models, such as U-Net, are prone to overconfident misclassifications on Out of Distribution (OOD) Data, posing severe risk of silent failures from a confident misclassification. Although methods like Deep Ensembles and Monte Carlo (MC) Dropout provide benchmarks for uncertainty quantification, their use during inference requires multiple forward passes increasing the computational overhead linearly making real time applications challenging. To address this gap, we introduce UnU-Net, a distance aware probabilistic architecture that incorporates a Spectral Normalized Neural Gaussian Process (SNGP) with the final pixel-wise prediction. By replacing conventional normalization techniques with Group Normalization and constraining the weights via Spectral Normalization, UnU-Net prevents feature collapse ensuring a preserved latent space distance. Evaluated on the Carvana and DeepCrack segmentation datasets, UnU-Net extracts precise segmentation masks and quantifies epistemic uncertainty flagging anomalies in a single forward pass. This architecture maintains the accuracy of the standard U-Net and outperforms conventional UQ methods while only requiring a single forward pass.

الكلمات المفتاحية

غير متوفر

التخصصات

رسائل وأطاريح مشابهة