Synergistic Perception and Generative Recomposition: A Multi-Agent Orchestration for Expert-Level Building Inspection
摘要
建筑立面缺陷检查是结构健康监测和可持续城市维护的基础任务，但由于几何变异极大、在复杂背景上的低对比度，以及复合缺陷（如裂缝伴随碎石）的内在复杂性，仍面临巨大挑战。此类特征导致严重的像素不平衡和特征模糊，加之高质量像素级标注稀缺，阻碍了现有检测和分割模型的泛化。为填补这一空白，我们提出了将缺陷感知视为协作推理任务而非孤立识别的统一多主体框架\textit{FacadeFixer}。具体而言，\textit{FacadeFixer} 编排专门的检测和分割主体以处理多类型缺陷干扰，并与生成主体协同实现语义重组。该过程将复杂缺陷从噪声背景中解耦，并将其真实合成到多样化的干净纹理上，生成具有精确专家水平掩码的高保真增强数据。为支持此目标，我们引入了覆盖六个主要建筑类别的综合多任务数据集，包含像素级标注。大量实验表明，\textit{FacadeFixer}显著优于当前最先进（SOTA）基准。具体而言，它在捕捉像素级结构异常方面表现卓越，并凸显生成合成是基础设施检查中应对数据稀缺的稳健解决方案。我们的代码和数据集将公开提供。
引用
@article{arxiv.2603.20143,
title = {Synergistic Perception and Generative Recomposition: A Multi-Agent Orchestration for Expert-Level Building Inspection},
author = {Hui Zhong and Yichun Gao and Luyan Liu and Xusen Guo and Zhaonian Kuang and Qiming Zhang and Xinhu Zheng},
journal= {arXiv preprint arXiv:2603.20143},
year = {2026}
}
备注
We are withdrawing this article because we recently identified a major methodological error regarding the multi-agent orchestration setup described in Section 4.2. This issue significantly impacts the final conclusions drawn in the paper. We sincerely apologize for any confusion this may have caused