Deep Learning Artwork Style Prediction and Similarity Detection
- Igor Sorarrain Rebollar 1
- Manuel Graña 1
- 1 University of the Basque Country (UPV/EHU), Donostia, Spain
- José Manuel Ferrández Vicente (dir. congr.)
- José Ramón Alvarez Sánchez (dir. congr.)
- Félix de la Paz López (dir. congr.)
- Hojjat Adeli
Publisher: Springer Suiza
ISBN: 978-3-031-06527-9
Year of publication: 2022
Pages: 289-297
Type: Book chapter
Abstract
The point of departure of this work is the aim to predict artwork style. The paper presents results retraining some of the most popular deep learning models for image classification, i.e. ResNet-34, ResNet-50, VGG-16, DenseNet-121, and a CNN model made from scratch over a dataset extracted from WikiArt for a Kaggle competition. This dataset is composed of 103253 images, categorized into 136 different artwork styles. We select 20 art styles that have enough image samples to allow for network training, achieving accuracy comparable to state of the art results. Moreover, we observe that the structure of the confusion matrix reflects the conceptual relations between the artwork styles, hence points to an induced similarity measure between styles of artwork instances