Abstract: This paper focuses on the small-sample problem in welding defect detection and covers multiple types of data, including X-ray testing images, visual inspection images, ultrasonic testing signals, and ultrasonic images. From the two perspectives of dataset construction and algorithm optimization, this paper systematically reviews the research progress on small-sample learning in welding defect detection. At the dataset level, techniques such as geometric transformation augmentation, generative data augmentation, and image quality enhancement are summarized, which alleviate small-sample constraints through sample expansion and quality optimization. At the algorithm level, the applications of transfer learning and meta-learning are analyzed in detail. Transfer learning adapts pretrained parameters from large-scale general datasets to defect detection tasks, while meta-learning strengthens the model’s ability to rapidly adapt to small-sample scenarios by enabling it to “learn how to learn.” This review compares the application effects and limitations of different methods, pointing out that data augmentation techniques have difficulty enriching the semantic diversity of defects, and that transfer learning may be limited in performance when data quality is poor. Meanwhile, the excellent application potential of meta-learning in defect detection is summarized. Although meta-learning has been less studied and applied in welding defect detection, it shows considerable development potential. This paper provides a reference for subsequent research and engineering applications related to small-sample problems in welding defect detection.