Picasso, Matisse, or a Fake? Automated Analysis of Drawings at the Stroke Level for Attribution and Authentication

Ahmed Elgammal, Yan Kang, Milko Den LeeuwAAAI 2018 · DOI

Introduction

Authenticating art has traditionally relied on two pillars: the subjective "connoisseur’s eye" and expensive technical laboratory analysis. While forgers can often mimic composition and source period-accurate materials, replicating the fluid, spontaneous nature of a master's individual brush or pen strokes remains a significant challenge. This paper introduces an AI-driven methodology that attributes drawings to masters like Picasso and Matisse by analyzing the "unintentional signature" found at the stroke level.

Methodology: Stroke-Level Analysis

The core of the proposed system is a multi-stage pipeline designed to quantify the characteristics of individual lines. Unlike traditional methods that look at the whole image, this approach breaks drawings down into their constituent parts.

Stroke Segmentation

Because drawings often consist of complex, intersecting lines, the researchers developed a skeletonization approach. By representing the drawing as a network and treating segmentation as a graph problem, they identify individual strokes by finding paths that minimize "bending energy" at junctions.

Feature Extraction

The system utilizes a hybrid approach to feature extraction:

  • Hand-crafted Geometric Features: These include Fourier descriptors for global shape, reconstruction error profiles for smoothness, and histograms of curvature and thickness. These features are particularly effective at detecting the "inhibited" or hesitant strokes typical of forgeries.
  • Deep Learned Features: Gated Recurrent Units (GRUs) are used to analyze the "tone" and local geometry. By "walking" along a stroke and processing image patches, the network captures subtle variations in ink density and local shape that are difficult to define manually.
Analysis of individual strokes reveals the "unintentional signature" of an artist, capturing subtle variations in line quality.

Datasets and Experimental Setup

The study utilized a primary dataset of 297 digitized drawings and prints, yielding over 80,000 individual strokes. The collection focused on works by Pablo Picasso, Henri Matisse, and Egon Schiele. To test the system's robustness, the authors commissioned five artists to create a "Fake Dataset" of 83 drawings designed to imitate these masters.

The dataset encompassed a wide variety of subjects and styles from masters like Picasso and Matisse, providing a rigorous benchmark for the AI.

Key Results

The system demonstrated high efficacy in both attribution and authentication:

  • Attribution Accuracy: The system achieved over 80% accuracy in attributing drawings to the correct artist by aggregating stroke-level classifications.
  • Forgery Detection: In "across-technique" tests, the classifiers rejected 100% of the fake drawings. The AI successfully identified that the imitations lacked the spontaneous stroke profiles of the original masters.
  • Feature Comparison: While GRUs excelled at general attribution with large datasets, hand-crafted features proved superior for detecting fakes, as they were less susceptible to the superficial stylistic similarities mimicked by forgers.

Conclusion

By focusing on how a line is drawn rather than what is drawn, this research provides a scientific, cost-effective tool for art authentication. The "unintentional signature" of a master's stroke serves as a powerful biometric for art. Future work aims to extend this analysis to the complex, multi-layered brushwork found in oil paintings.

Summary updated

BibTeX
@inproceedings{elgammal2018picasso,
  author = {Ahmed Elgammal and Yan Kang and Milko Den Leeuw},
  title = {{Picasso, Matisse, or a Fake? Automated Analysis of Drawings at the Stroke Level for Attribution and Authentication}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11313},
}
Figure