An ancient mural classification model with shallow texture enhancement and cross-layer attention aggregation

Ancient mural images are often degraded by weathering, fading, noise, and localized damage, leading to weakened textures and incomplete details that hinder classification. To address these challenges, we propose an improved network, R-EfficientNet, for robust mural image classification under complex degradation conditions. In the shallow layers, a multi-scale residual (MSR) module enhances pattern and boundary representation through parallel convolutional branches. At the intermediate stage, a lightweight adaptive feature fusion (AFF) module dynamically integrates cross-layer semantic information. In the classification stage, an attention pooling mechanism enables the model to emphasize discriminative regional features. Together, these components improve recognition stability and feature representation. R-EfficientNet achieves accuracy, precision, recall, and F1 scores of 88.79%, 89.37%, 88.98%, and 89.13%, respectively, which is 4.47% higher accuracy than EfficientNetV2-S and outperforms several classical CNNs in accuracy, generalization, and stability, demonstrating strong potential for mural digital preservation and intelligent recognition.