Enhancing feature interaction for improved generalization in few-shot metal surface defect segmentation

Metal surface defect segmentation enables detailed inspections of industrial components, making it vital for industrial production and quality control. However, current deep convolutional neural network-based methods are often trained on large-scale annotated data and cannot be generalized to unseen defect types. This hinders their real-world applicability when defect categories are scarce and dynamic. To address these limitations, a novel few-shot segmentation framework named Enhancing Feature Interaction Network (EFINet) is designed to create a class-agnostic model with minimal training data and generalize to unseen defect categories. To mitigate the high intra-class variations between defect samples, the Prior-guided Bidirectional Feature Interaction module first performs optimized feature interaction to enhance feature representation using support-query correlations and dual prior masks. Then, to address significant scale variation within a single image, the Context-aware Attention-guided Feature Aggregation module effectively aggregates contextual and attentional information, improving the ability of the model to segment defects of varying sizes. Additionally, to alleviate information loss during global pooling, particularly for defect samples with complex characteristics, a Prototype-Loss Compensation module is introduced to generate a compensatory prototype and provide auxiliary loss. Extensive experiments on the FSSD-12 and Surface Defects-4ⁱ benchmarks demonstrate that EFINet achieves state-of-the-art performance in various experimental settings. The code will be available at https://github.com/Hai-Jun-Yan/EFINet.