Edge-Friendly NAS Framework for Mural Pattern Recognition via Structure-Aware Feature Fusion

Abstract

Ancient mural recognition faces unique challenges due to degradation, stylistic variations, and domain-specific symbolism. We propose a lightweight, edge-deployable neural architecture search (NAS) framework—SG-NAS-MPR—designed for accurate mural pattern recognition. Our framework integrates gated convolutions with frequency-domain fusion in a structure-aware module to enhance features under visual noise. A contrast-aware NAS strategy tailors compact backbones for real-time inference. Experiments on Dunhuang mural datasets show that our method surpasses existing CNN and NAS models in accuracy (93.4%) and F1-score (0.922), whereas reducing latency and model size. This work enables efficient and interpretable recognition in cultural heritage computing, supporting mobile museum applications and AR-based mural analysis.