Building Deformation Simulation and Recognition via Localized Geometric Transformation and Collaborative Statistical Metrics

Building deformation, for example, tilting, settlement, and uplift, threatens structural safety, yet its rapid assessment is challenged by scarce real samples, subtle features, background clutter, and a lack of automated grading criteria. This study presents a task-specific computer vision framework integrating ROI-confined geometric simulation with multi-metric image-based assessment. For simulation, a localized geometric transformation applies rotation and vertical translation exclusively to the region of interest (ROI) of the building while preserving the background, generating eight deformation types with randomly sampled transformation magnitudes. For recognition, an edge-density-based strategy determines the building ROI from the reference image, and the same coordinates are used to crop the simulated image before pixel-wise differencing. A collaborative trio of metrics (SSIM, PSNR, and the threshold-exceeding pixel ratio) classifies deformations into four severity grades, each metric addressing a distinct distortion facet to provide complementary descriptions of image change. Evaluation using four building scenes and their synthetic deformation samples illustrates the spatial difference responses and metric-based screening outcomes under prescribed image transformations. The proposed approach offers a non-contact, cost-effective, and interpretable solution for routine building inspections and long-term health monitoring.