<p>Deep learning has driven major advances in computer vision, enabling highly accurate models for tasks such as image classification, object detection, and segmentation. However, these models typically rely on large-scale, high-quality labeled datasets, which are often expensive, time-consuming, or impractical to obtain. In many real-world scenarios, data are scarce, noisy, imbalanced, or weakly labeled, posing significant challenges to conventional training approaches. This survey reviews recent progress in developing computer vision methods that perform effectively under such constrained data conditions. We categorize the main forms of data limitations and explore key strategies to address them, including data augmentation, semi-supervised learning, transfer learning, few-shot learning, and self-supervised learning. We further discuss benchmark datasets, evaluation protocols, and emerging applications where these methods are critical. Unlike prior isolated surveys, this work integrates data scarcity, noise, and imbalance into a unified framework—linking taxonomy, methodological advances, benchmarks, and ethical considerations—to guide future research on building robust and data-efficient vision systems.</p>