In the era of intelligent manufacturing, the rapid and accurate classification of strip steel surface defects is crucial. Deep learning typically relies on a large number of parameters and labeled samples to achieve outstanding performance. However, acquiring a sufficient number of defects in actual steel production poses challenges, and a high number of parameters can impact the real-time performance of defect classification. To tackle these issues, an Extending Inter-class Distance real-time network using positional orientation transformation for few-shot strip steel surface defect classification is proposed(called the EidNet). EidNet utilizes a fewer parameters neural network as the feature extractor, enabling quick model convergence. To overcome the potential limitations of the fewer parameters model in expressing features, EidNet employs a no learnable parametric technique to artificially extend the inter-class distance in the metric space, utilizing directional transformation of prototype positions to manually design the direction of extending, dispersing prototypes that are different from the query sample class as much as possible, thereby enhancing classification performance. The model uses the Euclidean distance as its classifier to maintain a low overall number of parameters. Experimental results demonstrate that EidNet significantly enhances the real-time performance of defect classification, striking a balance between real-time requirement and classification accuracy, while also exhibiting superior generalization capabilities.
13 August 2025 in IEEE Transactions on Instrumentation and Measurement
Rapid and accurate classification of surface defects in the strip steel production process contributes to reducing production costs in factories. In this article, a dual feature-integration network (DFINet) is constructed for faster and more pragmatic few-shot strip steel surface defect classification scenarios. The model utilizes a relatively simple convolutional neural network (CNN) as the backbone to improve real-time performance. Bidirectional long short-term memory networks are employed to extract cross-image features (CIFs) from the support set, which are, then, fused with the structural features extracted by the backbone to obtain aggregated features for updating deeper support set features. The positions of support samples are adjusted by randomly cropping them into several subimages and fusing the features of these subimages to further update the positions of support samples in the metric space. Leveraging the nonparametric structure and strong adaptability of Euclidean distance, the model employs it as a classifier to further enhance real-time and classification performance. Additionally, to simulate more pragmatic few-shot strip steel surface defect classification scenarios, a modification is made to the common N-way, K-shot training mode, by changing the number of samples in the support set from fixed-shot to Random-shot. Extensive experiments demonstrate that the proposed method can rapidly and accurately classify surface defects in strip steel, even in more challenging few-shot classification scenarios.