Few-Shot Aviation Materials Surface Defection Classification: From Non-Standardized Image Acquisitions

Defect classification in the aviation industry plays a crucial role in manufacturing. However, in non-standardized captured scenes, the object to be inspected and the background are often complex and variable, e.g., lighting, angle, occlusion, etc., which can affect the core features of the image data. Therefore, achieving accurate classification using a limited number of defect samples from products photographed in non-standardized environments has emerged as a challenging yet promising research focus. To address this issue, this study proposes a learning approach that can be trained using standardized industrial defect datasets to detect defects in few-shot, non-standardized industrial defect production. Specifically, this approach initially utilizes the knowledge acquired from a standardized pre-training dataset to crop the data within the query set. Subsequently, an embedding extraction is employed to obtain feature embedding and classification results. Our practice demonstrates a performance improvement of roughly 5% on non-standard datasets compared to few-shot classification methods.