Abstract
The popularity of machine learning algorithms produced numerous applications in computer vision in the past 10 years. One application is art authentication, which assures that a piece of art is created by an artist. The models produced by machine learning algorithms provide an objective measure to authenticate an artist to their artwork collection. This article discusses an experiment using the residual neural network machine learning algorithm. This experiment demonstrates how a computer can distinguish between 34 and 958 artists with various degrees of confidence.
TopBackground
With the popularity of digital image processing and supervised machine learning, Johnson et al. (2008) provides objective measures for determining Van Gogh’s artistic style. This work branched off into numerous research efforts analyzing artistic style thus providing a basis to mitigate the issues of missing provenance with a digital signature of an artist’s work. The success of Russakovsky et al. (2015) winning the ImageNet challenge pushed image classification to new performance standards. From an artist classification perspective, Viswanathan and Stanford (2017) and Dobbs, Benedict, and Ras (2021) build on the success of ImageNet winners by applying residual neural networks to increase the performance of artist classification using the WikiArt data set. Likewise, Mensink and van Gemert (2014) and van Noord, Hendriks, and Postma (2015) apply machine learning algorithms to increase the performance of artist classification using the Rijksmuseum data set. The Rijksmuseum is the national museum of the Netherlands. They tell the story of 800 years of Dutch history, from 1200 to now. In addition, they organize several exhibitions per year from their own collection and with (inter) national loans (Rijksmuseum, 2021). The focus of this chapter relates to the Rijksmuseum dataset and ResNet 101 algorithm. (Dobbs et al., 2021) discuss additional background related to the OmniArt and WikiArt datasets and lower performing algorithms.
Key Terms in this Chapter
Confusion Matrix: A table of actual and predicted classes such that the intersection of said classes statistically define the corresponding level of confusion.
Digital Image Processing: The process of applying computer algorithms to digital images to satisfy a visual related computing task.
Transfer Learning: The process of leveraging existing learning models to facilitate the initial learning of related new problems.
Art Authentication: The process of proving a piece of art to be created by an artist.
Supervised Machine Learning: A type of machine learning for which labeled input data is used to train a model to determine an output classification.
Deep Learning: A type of machine learning which leverages a layered pipeline of neural networks which progressively extract features from input data.
Machine Learning: The application of statistical computer algorithms such that inference is used rather than explicit instruction to accomplish a task.