Industrial surface defect detection is constrained by the scarcity of defective samples and by the insufficient capacity of current segmentation methods to precisely delineate defect boundaries. To address these challenges, we propose F2P-Net, a few-sample, highly precise industrial surface defect segmentation framework composed of three core modules. ViCNet (ViT and CNN collaborative encoder network) integrates a vision transformer backbone with an auxiliary convolutional branch to retain robust large-model priors while enhancing sensitivity to fine-scale textures and local irregularities. AFDec (automated geometric prompt and multi-scale feature fusion decoder) employs automated geometric prompts to localize potential defect regions and fuses hierarchical multi-scale features to improve boundary delineation and mask consistency. EVPT (edge-enhanced visual prompt tuning) is a fine-tuning module incorporating edge-explicit visual prompt to facilitate effective industrial domain adaptation of large vision models. The proposed method achieves considerable performance over existing full-data training approaches in metrics including mAP, Recall, and IoU using only 1.76 %∼3.06 % of training images across NEU_Seg, MT, KolektorSDD2, and DAGM2007 datasets. Under full-data training, it attains state-of-the-art segmentation accuracies with IoU scores of 86.03 %, 92.57 %, 78.77 %, and 82.55 %, respectively. The network provides a novel solution for industrial applications with few-sample, high-precision defect segmentation. Code is available at https://github.com/kerongYan/F2P-Net.

中文翻译：

---

F2P-Net：一种基于有限数据的工业表面缺陷分割的混合提示增强双分支协作网络

工业表面缺陷检测受到缺陷样品稀缺和当前分割方法精确划定缺陷边界的能力不足的限制。为了应对这些挑战，我们提出了 F2P-Net，这是一个由三个核心模块组成的少量样本、高精度的工业表面缺陷分割框架。ViCNet（ViT 和 CNN 协作编码器网络）将视觉转换器主干与辅助卷积分支集成在一起，以保留稳健的大模型先验，同时增强对精细纹理和局部不规则性的敏感性。AFDec（自动几何提示和多尺度特征融合解码器）采用自动几何提示来定位潜在缺陷区域，并融合分层多尺度特征，以改善边界描绘和掩模一致性。EVPT（edge-enhanced visual prompt tuning）是一个融合边缘显式视觉提示的微调模块，用于促进大视觉模型的有效工业领域适配。所提出的方法在 NEU_Seg、MT、KolektorSDD2 和 DAGM2007 数据集中仅使用 1.76 %∼3.06 % 的训练图像，在包括 mAP、Recall 和 IoU 在内的指标方面取得了优异的性能。在全数据训练下，它获得了最先进的分割精度，IoU 得分分别为 86.03 %、92.57 %、78.77 % 和 82.55 %。该网络为工业应用提供了一种新颖的解决方案，具有少量样本、高精度的缺陷分割。代码可在 https://github.com/kerongYan/F2P-Net 获得。