AI-Driven Computer Vision for Industrial Automation and Domain Adaptation

Computer vision (CV) and machine learning (ML) are foundational technologies in modern manufacturing and central enablers of Industry 4.0, driving improvements in production efficiency, precision, and system interoperability. Despite their potential, the deployment of CV and ML models in industrial environments is often constrained by the reliance on large-scale, high-quality annotated datasets. Such datasets are costly to acquire, time-intensive to label, and susceptible to annotation errors. Moreover, conventional learning paradigms typically assume that training and deployment data follow identical distributions (i.i.d. assumption), an assumption that is frequently violated in real-world settings. In practice, source data may be unavailable due to storage limitations, restricted accessibility, or privacy concerns, while distributional shifts, class imbalance, and noisy labels further exacerbate performance degradation. Unsupervised domain adaptation (UDA) seeks to transfer knowledge from a labeled source domain to an unlabeled target domain with different data distributions. However, adaptation does not inherently ensure performance gains: negative transfer occurs when aligning mismatched or ambiguous features causes the adapted model to perform worse on the target domain than a non-adapted baseline. This issue is particularly pronounced when adaptation focuses solely on matching global (marginal) feature distributions without preserving class-level structure. Many existing UDA approaches emphasize marginal feature alignment, which can inadvertently align samples from different classes across domains, leading to class confusion and degraded discriminability. Discriminative alignment, which enforces class-consistent and decision-boundary-aware feature alignment, is therefore critical for learning transferable representations that remain both domain-invariant and class-separable. To tackle these data-centric and distribution-shift challenges, this thesis presents a threefold framework. First, we leverage synthetic data generation to mitigate dependence on expensive source data collection and manual annotation while ensuring scalability and diversity. Second, we propose various UDA frameworks (i.e., Raw-Mix, FD-SSL and WCT-Enhanced AdaIN) that jointly enforce marginal and discriminative feature alignment between source and real-world target domains, thereby reducing the need for extensive target data collection and labeling. Third, we introduce parameterfree test-time adaptation (TTA) methods named Adapt Without Training (AWT) that enable on-the-fly model adaptation using only unlabeled target data, addressing practical scenarios where source data is inaccessible during deployment. Extensive experimental evaluations validate the effectiveness of the proposed methods across diverse domain adaptation settings, including standard benchmark datasets and challenging synthetic-to-real industrial scenarios. The results demonstrate the robustness, generalization capability, and practical relevance of the proposed approaches for real-world manufacturing applications.