Multi-Scale Feature Pyramid Network with Bidirectional Attention for Efficient Mural Image Classification

This study proposes a mural classification model based on DenseNet201 FPN, which integrates bidirectional attention mechanism (Bi CBM), dynamic focus distillation loss, and convex optimization regularization. By using multi-scale feature fusion (28 × 28 × 256, 14 × 14 × 512, 7 × 7 × 1024) to enhance texture perception ability, and utilizing bidirectional LSTM iterative optimization of channel and spatial attention weights, the recognition performance of low-frequency categories (such as clothing textures) is significantly improved. By combining the dynamic temperature distillation strategy (T=3 → 1) with the balanced teacher model (ResNeXt101) and real label supervision, the rare category F1 score was increased by 6.1%. On the self built mural dataset (2000 images/26 categories), an accuracy rate of 87.9% was achieved (an improvement of 3.7% compared to DenseNet201), and the inference speed of edge devices reached 63ms/frame (Jetson TX2/8.1W), providing an efficient solution for the digitization of cultural heritage in resource constrained environments.