MSc Thesis Defense: Amirreza Khoshbakht, UNCERTAINTY-AWARE OPEN-SET RECOGNITION UNDER DOMAIN, CATEGORY, AND MODALITY SHIFTS, Date & Time: 13 July, 2026 – 01:30 PM, Place: FENS L063

UNCERTAINTY-AWARE OPEN-SET RECOGNITION UNDER DOMAIN, CATEGORY, AND MODALITY SHIFTS 

Amirreza Khoshbakht
Data Science, MSc Thesis, 2026

Thesis Jury

     Prof. Erchan APTOULA (Thesis Advisor)

  Assoc. Prof. Hüseyin ÖZKAN

  Prof. Behçet Uğur TÖREYİN 

Date & Time: July 13th, 2026 – 01:30 PM

Place: FENS L063

Keywords : Open-Set Recognition, Domain Generalization, Uncertainty Quantification, Image Classification, Deep Learning

Abstract

The growing use of deep learning in image classification, medical imaging, and remote sensing has increased the need for reliable models that can operate under real-world uncertainty. However, conventional recognition systems often assume that test samples belong to known classes and follow the same distribution as the training data, leading to degraded performance under unseen categories, domain shifts, and modality mismatches. To tackle these challenges, five complementary approaches are investigated. First, a semantically guided gradient matching framework for open-set domain generalization is proposed, which incorporates class-level semantic relationships into dualistic meta-learning to improve decision boundaries under domain and category shifts. Second, a hierarchical uncertainty-aware deep simplex classification model is introduced for open-set medical image recognition, combining multi-scale features, Simplex ETF classifiers, and uncertainty fusion. Third, an open-set domain generalization framework for hyperspectral image classification is developed, integrating frequency-domain disentanglement, spectral–spatial feature extraction, evidential uncertainty quantification, and adaptive pathway selection. Fourth, a hyperbolic prototype learning method is proposed for open-set joint HSI–LiDAR classification, using Poincaré ball representations and counterfactual modal mismatch detection to exploit spectral–elevation consistency. Finally, an adaptive contrastive semantic reconstruction framework is presented for hyperspectral open-set recognition, combining cross-group spectral interaction, contrastive latent regularization, Fisher-calibrated score fusion, and automatic threshold calibration. Experimental results across multiple benchmarks demonstrate the effectiveness of the proposed approaches in improving robustness, uncertainty awareness, and generalization for open-set recognition under diverse domain, category, and modality shifts.