Classification of Geometric Forms in Mosaics Using Deep Neural Network.

Area of Science:

• Computer Vision
• Digital Art History
• Machine Learning

Background:

• Image processing in fine arts presents challenges in analyzing complex geometric patterns.
• Digital reconstruction and photogrammetry are increasingly used for artwork analysis.
• Automated classification of geometric forms in artworks requires sophisticated computational methods.

Purpose of the Study:

• To develop and evaluate a deep learning-based technique for classifying geometric forms in artworks.
• To demonstrate the effectiveness of a convolutional neural network (CNN) framework for art pattern recognition.
• To test the robustness of the proposed method on a case study involving a Roman mosaic.

Main Methods:

• A convolutional neural network (CNN) framework was designed, incorporating convolution, pooling, and dense layers.
• A dataset of geometric forms (triangles, squares, circles, octagons, leaves) was created from original and rectified images of a Roman mosaic.
• Close-range photogrammetry was used for the digital reconstruction and orthophoto generation of the mosaic.

Main Results:

• The proposed CNN-based method achieved a classification accuracy exceeding 97% for geometric forms in the Roman mosaic.
• The framework demonstrated robustness in analyzing partially deformed geometric shapes.
• Performance comparison with standard deep learning frameworks indicated the efficacy of the proposed approach.

Conclusions:

• The developed deep learning technique is highly effective and robust for analyzing geometric forms in artworks.
• The method shows significant potential for application in broader pattern identification problems within art analysis and digital heritage.
• This approach advances the integration of AI in the study and preservation of cultural heritage.