Geometry-Aware Feature Enhancement With Linear Attention for Robust UAV Photogrammetric Reconstruction Under Weak and Isomorphic Textures

UAV photogrammetry constitutes a fundamental technology for the lifecycle maintenance and digital twin construction of large-scale transport infrastructure. However, standard Structure-from-Motion (SfM) pipelines frequently falter in scenarios characterized by weak textures, such as asphalt, and repetitive patterns. This deficiency leads to severe feature ambiguity and sparse reconstruction voids in large-scale infrastructure scenes. Furthermore, existing deep learning descriptors typically neglect explicit spatial attributes and suffer from the computational burden of quadratic-complexity attention mechanisms, hindering deployment on edge devices. Building on the FeatureBooster-style descriptor enhancement paradigm, this study adapts a lightweight geometry-aware reconstruction framework to unmanned aerial vehicle infrastructure inspection. The methodology integrates a dual-stream descriptor enhancement model. Following the descriptor-boosting idea of combining local descriptors with geometric keypoint attributes, the adapted model embeds spatial attributes into the feature space to alleviate isomorphic ambiguity. Meanwhile, the modified cross-perception module replaces the original Attention-Free Transformer (AFT) -Simple setting with AFT-Full and incorporates SwiGLU to capture global context efficiently. Experiments on real-world datasets of complex interchanges, urban highways, and campus scenes validate that the adapted descriptor enhancement strategy effectively reduces point cloud voids and reduces reprojection error by approximately 17.3% compared with the original SIFT baseline on Dataset 1. Notably, this study empirically identifies an efficiency compensation phenomenon, wherein superior feature quality accelerates downstream geometric verification and optimization stages. Consequently, although feature enhancement introduces marginal overhead, the overall reconstruction time is reduced in certain datasets. This work provides an application-oriented adaptation and evaluation of FeatureBooster-style descriptor enhancement for geometry-consistent and computationally efficient infrastructure digitalization.