Few-shot defect recognition of metal surfaces via attention-embedding and self-supervised learning

Vision-based defect recognition in the metal industry is an important and challenging task. Many mainstream methods based on supervised learning need a large amount of annotated defect data. However, the data acquisition in industrial scenarios is quite difficult and time-consuming. To alleviate the problem, in this paper, we propose a few-shot defect recognition (FSDR) method for metal surfaces by attention-embedding and self-supervised learning. The proposed method includes two stages,pre-training and meta learning. In the pre-training stage, an attention embedding network (AEN) is designed for better learning the defect local correlation of multi-receptive field and reducing the background interference, and to train a robust AEN without any annotations, a multi-resolution cropping self-supervised method (MCS) is developed for better generalizing the few-shot recognition task. Then, in the meta learning stage, to generate the embedding features, the defect images are encoded by the pre-trained AEN, and we also design a query-guided weight (QGW) to address the bias of embedding feature vector. The classification information is gained by computing the distance of the embedding feature vector of each category. We evaluate the proposed FSDR on the NEU dataset and the experiments show competitive results in 1-shot and 5-shot defect recognition tasks compared with the mainstream methods.