Dual-Path Detection and Physical Simulation for Sustainable Crack Evolution Monitoring in Ancient Chinese Murals

This study presents a dual-path framework for mural heritage conservation that integrates lightweight crack detection with physics-based simulation. In the detection path, a YOLO-MME model based on a MobileNetV4 backbone, enhanced multi-head cross-attention, and an improved IoU-based loss achieves an effective balance between accuracy and computational efficiency for resource-limited deployment. In the simulation path, a bi-layer RFPA3D model is employed to analyse the effects of loading ratio (λ=Δy/Δx) and overlay thickness (t) on crack evolution across four stages: initiation, propagation, coalescence, and saturation. Validation on 20 paired mural samples demonstrates strong cross-sectional consistency with field observations, quantified using a multi-level correspondence marker system, yielding high skeleton overlap (IoU = 0.82) and reduced geometric deviation relative to uncalibrated baselines. Importantly, the evaluation framework supports both agreement assessment and failure-mode diagnosis. By coupling data-driven detection with physical modelling, the framework enables standardized documentation and mechanism-oriented conservation decision-making.