FH-Net: Frequency-domain and hierarchical dual enhancement for few-shot surface defect segmentation

Abstract

In real-world industrial manufacturing, surface defect segmentation is severely challenged by the extreme scarcity of annotated defect samples, which renders exhaustively supervised approaches impractical. Although meta-learning based few-shot segmentation paradigms have made initial progress, existing methods still suffer from two prominent limitations. First, the limited number of support set samples leads to insufficient feature diversity, rendering the model prone to overfitting during training. Second, feature extraction restricted to a single hierarchical level and lacking contextual guidance constrains the representational capacity for complex and multi-scale defects. To address these issues, a novel few-shot surface defect segmentation method termed FH-Net is proposed, which aims to achieve high-precision segmentation of previously unseen categories using only a limited number of annotated samples. Specifically, a Semantic-Preserving Frequency Perturbation (SPFP) module is designed, which enriches support set feature diversity through foreground- and global-level amplitude perturbations in the Fourier domain, thereby mitigating the risk of overfitting under limited supervision. Furthermore, a Hierarchical Context Aggregation Module (HCAM) is constructed, which strengthens multi-scale defect representation via parallel feature extraction branches coupled with an attention-driven adaptive weighting fusion mechanism. Extensive experiments on FSSD-12 and Surface Defects-4i demonstrate that FH-Net achieves competitive segmentation performance compared with representative recent methods under both the 1-shot and 5-shot settings. Ablation studies further confirm the contribution of each core module to the overall performance improvement.