Jumping into the artistic deep end: building the catalogue raisonné

Abstract

The catalogue raisonné compiled by art scholars holds information about an artist’s work such as a painting’s image, medium, provenance, and title. The catalogue raisonné as a tangible asset suffers from the challenges of art authentication and impermanence. As the catalogue raisonné is born digital, the impermanence challenge abates, but the authentication challenge persists. With the popularity of artificial intelligence and its deep learning architectures of computer vision, we propose to address the authentication challenge by creating a new artefact for the digital catalogue raisonné: a digital classification model. This digital classification model will help art scholars with new artwork claims via a tool that authenticates a proposed artwork with an artist. We create this tool by training a machine learning model with 90 artists having at least 150 artworks and achieve an accuracy of 87.31%. We use the ResNet Convolutional Neural Network to improve accuracy and number of artist classes over state-of-the-art artist classification experiments using the WikiArt database. We address inconsistencies in the way scholars approach artist classification by providing a consistent method to recreate our dataset and providing a consistent method to calculate performance metrics based on imbalanced data.

Contemporary Art Authentication with Large-Scale Classification

“…In Table 3, accuracy results are compared using the same macro-balanced accuracy metric. In all cases, the Artfinder experiments outperform WikiArt and Rijksmuseum experiments 10+% [3,4]. This increase in accuracy for contemporary art may be because experiments start with over twice as many artists, and the annealing process can select the best artists for classification once reaching a comparable artist count in previous experiments.…”

“…In Table 3, the artist count is inversely proportional to validation and test accuracy. Given the state-of-the-art research using multi-classification for image-only art authentication, this behavior is expected [3,4,23]. With these experiments, the goal is to reproduce the counterintuitive phenomenon wherein many classes can improve multi-classification metrics as the number of classes grows [24].…”

“…In the past five years, art authentication received increased attention due to artificial intelligence, digital image processing, forensic techniques, and legal cases. From an artificial intelligence perspective, supervised deep learning algorithms, when applied to images of paintings, have attained an accuracy of 67.78% in authenticating art for 90 artists using the WikiArt dataset [3], and an accuracy of 32.40% in authenticating art for 1199 artists using the Rijksmuseum dataset [4]. On the digital image processing front, an accuracy of 91.7% was achieved in authenticating art for two artists using the principal component analysis (PCA) and a custom van Gogh and Raphael dataset.…”