Uncertainty-aware Deep Learning Methods for Image Classification, Object Detection and Segmentation

Deep learning has garnered significant attention over the last decade, emerging as a transformative force in science and technology. As the foundation of modern artificial intelligence (AI), deep learning models have revolutionized computer vision, impacting numerous areas such as facial recognition, autonomous driving, medical image segmentation, and generative AI systems. These advancements have profoundly changed daily life, demonstrating the immense potential and versatility of AI technologies.

Despite the rapid progress and widespread adoption of deep learning technologies, critical challenges persist. Modern deep learning models often suffer from overconfidence, susceptibility to distributional shifts, and vulnerabilities to adversarial attacks. These limitations raise serious concerns about the reliability and trustworthiness of AI systems, particularly in safety-critical applications where erroneous predictions can have life-threatening consequences. As a result, building reliable, explainable, and uncertainty-aware AI systems has become a focal point of research.

This thesis addresses pressing challenges in large-scale computer vision by focusing on both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). While these models have achieved significant advances in performance, accuracy, and scalability, ensuring their reliability under distribution shifts remains a fundamental barrier for deployment in real-world and safety-critical scenarios. To address this, the thesis emphasizes the quantification and management of uncertainty as a means to achieve robustness, safety, and trustworthiness in deep learning models. Specifically, this work develops practical and scalable uncertainty quantification methods, including a novel Bayesian object detection framework that leverages Gaussian weight sampling from pre-trained networks for effective out-of-distribution (OOD) detection without incurring high computational costs or requiring synthetic data. Experimental results demonstrate substantial improvements, including up to an 8.19\% reduction in FPR95 and a 13.94\% increase in AUROC, thus enhancing the reliability of object detection in open-world settings. For Vision Transformers, the Prior-augmented Vision Transformer (PViT) is introduced, which utilizes prior knowledge from pre-trained models to robustly separate in- and out-of-distribution samples by quantifying the divergence in predicted class logits, outperforming state-of-the-art OOD methods across the ImageNet benchmark and seven additional OOD datasets. Furthermore, this thesis contributes a new benchmark dataset for complex cell morphology segmentation and proposes an uncertainty-aware framework that incorporates virtual outlier sampling, leading to up to a 7.74\% increase in Dice Similarity Coefficient and significant reductions in boundary errors. Collectively, these contributions advance the interpretability, reliability, and deployment-readiness of computer vision models in complex, real-world scenarios. The public release of codebases and datasets further amplifies the scientific and societal impact, paving the way toward safer and more dependable AI systems.