AI-POWERED INSIGHTS

Innovative Hybrid Architecture

The proposed method integrates the strengths of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). CNNs are leveraged for their excellence in local feature extraction, identifying fine-grained details like stroke thickness and angles using convolution operations (as per Equation 1).

ViTs are utilized to process image patches, enabling the model to capture long-range dependencies and global contextual information across an image (as per Equation 2). This overcomes the inherent limitation of CNNs in understanding global context.

The core of the innovation lies in the CVIT-CFA architecture, an enhanced MobileViT-based model. This hybrid approach combines CNNs for local feature extraction with transformers for global attention. A crucial component is the Calligraphic Feature Attention (CFA) module, specifically designed to enhance sensitivity to distinct stroke features—such as direction, thickness, and spacing—which are vital for distinguishing between different calligraphers. The CFA module refines feature maps at multiple stages of the network.

The CFA module's structure involves global average pooling, followed by a 1x1 convolution with Group Normalization and Mish activation to reduce dimensionality. Horizontal and vertical convolutions are then applied to capture distinct structural features. Another 1x1 convolution, GN, and Mish further process features. Finally, feature fusion with a sigmoid activation function generates an attention map, which is multiplied element-wise with the original feature map to enhance relevant features.

Benchmark Performance and Key Findings

For the experiments, a custom dataset was meticulously constructed comprising over 6700 samples from five renowned historical Chinese calligraphers: Huang Tingjian (1340 samples), Liu Gongquan (1351 samples), Song Huizong (1353 samples), Wang Xizhi (1343 samples), and Yan Zhenqing (1351 samples). All images were cropped to 64x64 pixels. The experimental setup included a batch size of 32, input image size 64x64, AdamW optimizer with a learning rate of 1x10-4, and training for 100 epochs. The dataset was split 80% for training, 10% for validation, and 10% for testing.

The performance of the proposed architecture was evaluated using standard metrics: accuracy, precision, recall, and F1 score. Comparative analysis against classic models—ShuffleNet, DenseNet, EfficientNet, and MobileViT—demonstrated the superior performance of our method.

Key Findings:

• The proposed hybrid model achieved a remarkable 97.4% accuracy, significantly outperforming all other methods tested.
• Analysis of loss curves showed that our method maintained the lowest final loss, indicating robust and effective learning.
• While EfficientNet achieved a lower loss after ~20 epochs, its final accuracy was still surpassed by our proposed model. MobileViT improved steadily to ~91% accuracy, but also fell short of the hybrid model's performance.
• The confusion matrix revealed high accuracy per category (132-135 correct classifications) and minimal, evenly distributed misclassifications (1-3 per category), indicating strong generalization across different calligraphers.

Strategic Implications for Cultural Heritage

The proposed hybrid CNN-ViT architecture, enhanced by the Calligraphic Feature Attention (CFA) module, effectively captures both local and global features essential for distinguishing intricate variations in calligraphic styles. This research provides a novel, quantifiable method for analyzing and distinguishing personal styles in historical Chinese calligraphy.

Key Applications:

• Art Collection and Authentication: Provides a robust, objective tool for verifying the authorship and authenticity of ancient Chinese calligraphy pieces, reducing reliance on subjective human expertise.
• Academic Research: Offers new avenues for cultural heritage studies, enabling deeper analysis of calligraphic evolution and individual styles.
• Digital Archiving: Enhances the accuracy and detail of digital archives for historical documents, ensuring better preservation and accessibility.
• Automated Style Recognition: Laying the groundwork for automated systems that can assist practitioners in learning and practicing different styles by providing objective evaluation.

This work significantly contributes to the intersection of artificial intelligence and cultural heritage, opening new possibilities for safeguarding and understanding invaluable artistic traditions.