Scholarlink

<p>Mural image recognition is a core task in the digital preservation of cultural heritage, as murals embody substantial historical and artistic value. However, acquisition and restoration are often constrained by scarce samples, missing information, and intricate fine-grained details. To address these challenges, we propose an SPD-Conv–augmented mural image classification method built on an EfficientNetV2 backbone. A multi-scale Feature Pyramid Network (FPN) is employed to enhance cross-scale feature fusion, and a bidirectional Convolutional Block Attention Module (Bi-CBAM) is introduced to emphasize salient regions via two-way feature interaction. In addition, a dynamic knowledge distillation strategy is adopted to balance model compactness with accuracy. The proposed approach is particularly effective at preserving high-frequency textures (e.g., bodhisattva drapery patterns and animal fur) and strengthening representations in edge-attenuated regions.On an in-house mural dataset comprising 5,390 images, our method achieves 87.9% accuracy, representing a 3.7 percentage-point improvement over DenseNet201. Ablation studies show that FPN and Bi-CBAM contribute + 0.9 and + 1.2 percentage points to accuracy, respectively. Compared with nine mainstream baselines—including Swin Transformer and ConvNeXt—our method attains the best mAP (85.7%) and the fastest inference speed (23 ms per frame). These results demonstrate that the proposed framework substantially improves inference efficiency while maintaining high accuracy, offering a practical solution for efficient mural recognition and the digital safeguarding of cultural heritage.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mural image classification and lightweight feature fusion design based on the SPD-Conv module

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Shulan Wang,
  • Siyu Liu

摘要

Mural image recognition is a core task in the digital preservation of cultural heritage, as murals embody substantial historical and artistic value. However, acquisition and restoration are often constrained by scarce samples, missing information, and intricate fine-grained details. To address these challenges, we propose an SPD-Conv–augmented mural image classification method built on an EfficientNetV2 backbone. A multi-scale Feature Pyramid Network (FPN) is employed to enhance cross-scale feature fusion, and a bidirectional Convolutional Block Attention Module (Bi-CBAM) is introduced to emphasize salient regions via two-way feature interaction. In addition, a dynamic knowledge distillation strategy is adopted to balance model compactness with accuracy. The proposed approach is particularly effective at preserving high-frequency textures (e.g., bodhisattva drapery patterns and animal fur) and strengthening representations in edge-attenuated regions.On an in-house mural dataset comprising 5,390 images, our method achieves 87.9% accuracy, representing a 3.7 percentage-point improvement over DenseNet201. Ablation studies show that FPN and Bi-CBAM contribute + 0.9 and + 1.2 percentage points to accuracy, respectively. Compared with nine mainstream baselines—including Swin Transformer and ConvNeXt—our method attains the best mAP (85.7%) and the fastest inference speed (23 ms per frame). These results demonstrate that the proposed framework substantially improves inference efficiency while maintaining high accuracy, offering a practical solution for efficient mural recognition and the digital safeguarding of cultural heritage.