Scholarlink

<p>The automatic recognition of artistic styles is an important part of the management, analysis, and access to the huge digital cultural heritage archives. Nevertheless, the aesthetic complexity and minor distinctions between works of art become serious obstacles to this process. To overcome this, this study introduces a multi-stage, novel technique that, in comparison to current techniques, intelligently derives and combines local and global features to obtain a more comprehensive style representation. In the suggested approach, initially, the image is split into visual parts with the help of the K-means algorithm, and the local characteristics of each part are obtained with the help of specific CNNs. Simultaneously, a more detailed CNN obtains global features of the whole picture. The two sets of features are then intelligently fused with an attention mechanism. In the third step, an Autoencoder (AE) is employed to downsize the dimensions of the extracted features and eliminate noise and redundant data because of the large dimensionality of the extracted features. Lastly, a SoftMax classifier is used to perform classification. The proposed method was assessed using the Metropolitan Museum of Art (MET) Public Archives dataset in this study. Our approach on the MET dataset had an 89.07% accuracy and F-measure of 0.8867, which proves it to be superior.</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

Artistic style classification in traditional cultural heritage using convolutional neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Bo Xu,
  • Xiaomeng Lu

摘要

The automatic recognition of artistic styles is an important part of the management, analysis, and access to the huge digital cultural heritage archives. Nevertheless, the aesthetic complexity and minor distinctions between works of art become serious obstacles to this process. To overcome this, this study introduces a multi-stage, novel technique that, in comparison to current techniques, intelligently derives and combines local and global features to obtain a more comprehensive style representation. In the suggested approach, initially, the image is split into visual parts with the help of the K-means algorithm, and the local characteristics of each part are obtained with the help of specific CNNs. Simultaneously, a more detailed CNN obtains global features of the whole picture. The two sets of features are then intelligently fused with an attention mechanism. In the third step, an Autoencoder (AE) is employed to downsize the dimensions of the extracted features and eliminate noise and redundant data because of the large dimensionality of the extracted features. Lastly, a SoftMax classifier is used to perform classification. The proposed method was assessed using the Metropolitan Museum of Art (MET) Public Archives dataset in this study. Our approach on the MET dataset had an 89.07% accuracy and F-measure of 0.8867, which proves it to be superior.