Painting authentication using CNNs and sliding window feature extraction.

Area of Science:

• Art authentication
• Computer vision
• Machine learning

Background:

• Painting authentication is complex, combining connoisseurship and technical analysis.
• Data scarcity poses challenges for computational authentication methods.
• Stylistic analysis requires focusing on artistic execution over subject matter.

Purpose of the Study:

• To authenticate a painting attributed to Paolo Veronese using a novel CNN approach.
• To address severe data scarcity in art authentication through a patch-based strategy.
• To analyze microtextural and brushstroke features for stylistic comparison.

Main Methods:

• Developed a custom shallow convolutional neural network (CNN) for multichannel image inputs (RGB, grayscale, edge maps).
• Employed a sliding-window strategy to extract overlapping patches, expanding the dataset.
• Utilized regularization techniques for generalization and painting-level cross-validation to prevent data leakage.

Main Results:

• Achieved high classification performance: 94.51% accuracy and 0.99 Area Under the Curve (AUC).
• Generated probability heatmaps highlighting stylistic coherence in authentic works and fragmentation in others.
• The painting under examination showed partial stylistic affinity (61% probability) with authentic Veronese works.

Conclusions:

• Patch-based CNN models are effective for stylistic analysis in art authentication, particularly with limited data.
• The model provides probabilistic evidence of stylistic affinity, complementing traditional authentication methods.
• Definitive attribution requires integrating AI-driven stylistic analysis with historical, technical, and provenance data.