Abstract: To address the depth estimation confusion problem caused by the lack of spatial correlation mechanisms and local feature decoupling capabilities in existing methods for ancient architecture scenes with interleaved spatial hierarchies and dense edges, an unsupervised monocular depth estimation method based on local centripetal deduction was proposed. First, an encoder was adopted to extract multi-scale features from the input image. A local centripetal deduction strategy was constructed to reshape the spatial structure of features by utilizing the spatial correlation between the central pixel and neighboring pixels within a local attention window. In addition, a feature decoupling strategy based on a gated screening mechanism was introduced to reconstruct the encoder features, thereby effectively alleviating the depth estimation confusion problem. Then, by integrating the two aforementioned strategies, a depth estimation network based on local centripetal deduction was constructed to achieve the depth parsing of complex spatial hierarchies and the fine preservation of dense edges in ancient architecture scenes. Finally, under the unsupervised learning framework, the minimum reprojection loss and the edge-aware smoothness loss were adopted to realize the end-to-end joint optimization of the method. The experimental results on a self-built ancient architecture dataset show that, compared with Monodepth2, which performs the best among the comparison methods, the absolute relative error of the proposed method is reduced by 4.86%, and the accuracy (threshold $\delta < 1.25$) is increased by 0.64 percentage points. The method effectively overcomes the depth feature confusion problem brought by interleaved structures and dense edges, and significantly improves the three-dimensional (3D) perception capability for complex geometric details of ancient architectures. It provides precise geometric priors and technical support for the component-level high-fidelity mapping, non-destructive spatial analysis, and 3D digital archiving of ancient architectures.

Key words: ancient architectures, monocular depth estimation, unsupervised learning, local centripetal deduction, depth confusion

摘要： 针对现有方法在空间层次交错且边缘密集的古建筑场景中，因缺乏空间关联机制与局部特征解耦能力而导致深度估计混淆的问题，提出一种基于局部向心推演的无监督单目深度估计方法。首先，采用编码器提取输入图像的多尺度特征，利用局部注意力窗口内中心像素与邻域像素间的空间关联，构建局部向心推演策略以重塑特征空间结构，并引入基于门控筛选机制的特征解耦策略对编码器特征进行重构，从而有效缓解深度估计的混淆问题；其次，融合上述两种策略，构建一个基于局部向心推演的深度估计网络，实现对古建筑场景复杂空间层次的深度解析与密集边缘的精细保持；最后，在无监督学习框架下，采用最小重投影损失与边缘感知平滑度损失实现该方法的端到端联合优化。在自建古建筑数据集上的实验结果表明，所提方法相较于对比方法中表现最优的Monodepth2，绝对相对误差降低了4.86%，准确率（阈值δ<1.25）提高了0.64个百分点。该方法有效克服了因结构交错与边缘密集带来的深度特征混淆难题，显著提升了对古建筑复杂几何细节的三维感知能力，为古建筑的构件级高保真测绘、无损空间分析及三维数字化存档提供了精确的几何先验与技术支撑。

关键词: 古建筑, 单目深度估计, 无监督学习, 局部向心推演, 深度混淆