Designing an edge-friendly hybrid vision model for architectural health monitoring and crack measurement

Vision-based structural health monitoring (SHM) heavily relies on accurate crack segmentation. However, deploying heavy-weight models on edge devices is hindered by immense computational demands, while traditional lightweight models suffer from severe false-negative rates when detecting micro-cracks. To address this algorithmic dilemma, we propose an edge-friendly, high-recall hybrid segmentation architecture. The model employs MobileViT-v2 to extract robust global-local features, utilizing an Atrous Spatial Pyramid Pooling (ASPP) module and a four-stage cascade decoder to explicitly reconstruct high-fidelity crack edges. To mitigate missed detections, a cost-sensitive FocusCrackLoss is tailored to asymmetrically penalize hard-to-detect micro-cracks. Evaluated on the CRACK500 dataset, our model achieves a safety-critical recall of 91.91% and a Crack Intersection over Union (IoU) of 62.35%, deliberately prioritizing the minimization of undetected structural defects. Notably, the model requires only 6.30 M parameters and 10.45 G floating-point operations (FLOPs)—representing less than 1/6 and 1/16 of the computational overhead of standard heavy-weight benchmarks, respectively—indicating a computational profile suitable for subsequent evaluation on embedded graphics processing unit (GPU)-based Edge-AI platforms. To evaluate practical utility, we implemented a reference-object-free, scale-dependent physical quantification pipeline equipped with a 30-mm morphological filter. Cross-dataset evaluations on real-world bridge-deck images demonstrate that the pipeline can autonomously extract single-pixel skeletons, provide scale-dependent width estimates, and filter broad spalling-like false-positive regions, supporting automated SHM under complex surface conditions.