A real-time network using Feature Augmentation Recombination for few-shot strip steel surface defect classification

In the production process of strip steel, the occurrence of production defects is inevitable, timely detection of surface defects is of great significance for improving product performance and corporate economic benefits. Therefore, the classification of strip steel defects has become one of the research topics in deep learning. Traditional deep learning classification methods rely heavily on large amounts of training data and numerous training parameters. However, in the actual production process of strip steel, obtaining samples and effectively labeling them becomes increasingly challenging. As a result, it is difficult for traditional deep learning methods to achieve rapid and accurate defect classification under the condition of limited samples. In this paper, a real-time network using Feature Augmentation Recombination (FAR) for few-shot strip steel surface defect classification is introduced. The network model is capable of meeting the current industrial production requirements for high precision and high real-time performance. FARNet employs a lightweight network EfficientNet, to enhance the model’s real-time performance. The FAR module can augment and reorganize the feature representations of defect samples in the query set during the training phase. This enables the model to learn better feature extraction parameters for classification tasks, thereby improving its generalization ability. The Euclidean distance is adopted as the evaluation metric, which further reduces the number of training parameters and enhances the model’s real-time performance. Extensive experiments have demonstrated the model’s high precision and real-time performance, as well as its certain generalization capability.