Evaluation of Few-Shot Learning Algorithms, Training Methods, Backbones and Learning Task for Crack Detection in Manufacturing

Ensuring the structural integrity of components in manufacturing is paramount, with crack detection being a critical aspect of quality control. Traditional methods, which rely on either extensive labeled datasets or manual inspection, are both time-consuming and prone to human error. This study explores the application of few-shot learning (FSL) for crack detection to address these issues by utilizing minimal labeled data to train robust models. We compare the performance of models using episodic learning and transfer learning, along with inductive and transductive algorithms, using a Convolutional Neural Network (CNN) backbone for tasks involving crack detection and pseudo training. Our results show that transfer learning outperforms episodic learning, using novel data for validation during training producing the best-performing models. Furthermore, our findings suggest that limited labeled data is more effective than a large amount of unlabeled data.