VPGFNet: Cross-domain few-shot visual prompt graph fusion network for industrial defect segmentation

Abstract In industrial visual scenarios, challenges such as limited labelled samples, complex defect patterns, and significant distribution differences between training and test data are common. Existing few-shot semantic segmentation (FSS) methods primarily focus on in-domain tasks, but they often fail to maintain stability and generalization in cross-domain industrial settings, where the training and test data originate from different domains. In order to address this issue, we propose a cross-domain few-shot semantic segmentation (CD-FSS) framework named visual prompt graph fusion network (VPGFNet). VPGFNet uses the model’s own predicted masks as guidance and introduces three key modules. First, the grid-based visual prompting module enhances local structural perception. Second, the graph structure interaction module captures global semantic consistency among visual prompts. Third, the bidirectional optimization prediction module improves alignment between support and query features. VPGFNet eliminates the reliance on external segmentors by using self-guided prompts, and shows better adaptability in domain-shifted settings. Experiments on three industrial defect datasets (SSD-12, Surface Defects-4i, and ESDIs-SOD) show that our method outperforms state-of-the-art models by $4.8\%$ and $5.1\%$ in 1-shot and 5-shot settings, respectively.