Research and Implementation of Mural Classification Based on Lightweight Network

Dunhuang murals are a crucial historical and cultural heritage, presenting challenges in feature extraction and classification due to their number, age, and similarity. This study introduces SER-Net, a lightweight classification network for real-time mural analysis on mobile devices. A dataset covering nine dynasties—Early Tang, Northern Wei, Northern Zhou, Peak Tang, Sui, Late Tang, Middle Tang, Five Dynasties, and Western Wei—was created through manual collection and annotation. Data augmentation addressed uneven image distribution. SER-Net, based on RepVGG and ResNet18, features the SE D-Block module, which integrates SE attention and Channel-Shuffle mechanisms for enhanced feature fusion while using deep separable convolution to control model size. Experimental results show SER-Net reduces model size, boosts efficiency, and increases accuracy.