Fig. 2
Image representations in the complexity-entropy (C-H) plane and pipelines of image similarity measures through different image representation spaces. (a) Three reputable exemplary masterpieces chosen and localized in the C-H plane according to the given embedding parameters (\(\:{d}_{x}={d}_{y}=2\)), to aid viewers’ understandings of the image representations the plane encompasses. (b) A schematic diagram of an image similarity computation pipeline through a neural network. We measure pairwise cosine similarities of a 200-dimensional embedding vector including both low- and high-level feature maps extracted from individual visual artworks. (c) A schematic diagram of the image similarity computation pipeline through the SIFT descriptor. We measure pairwise Jaccard similarities of the SIFT descriptor matching a 128-dimensional vector of low-level features extracted from individual visual artworks. In addition to the image representations the C-H plane encompasses, we adopt and use two different types of image processing methods to obtain multi-level image features and measure their similarities: the ResNet architecture and the SIFT algorithm. Through our approach, we investigate characteristics of image representations in the C-H plane with image similarity measures that aggregate both low- and high-level features. The implemented images (1–3) are “Composition 2” by Piet Mondrian, 1929 (public domain, retrieved from WikiArt, https://www.wikiart.org/en/piet-mondrian/composition-2), “Rhythm” by Robert Delaunay, 1912 (public domain, retrieved from WikiArt, https://www.wikiart.org/en/robert-delaunay/rhythm-1), and “The Starry Night” by Vincent van Gogh, 1889 (public domain, retrieved from WikiArt, https://www.wikiart.org/en/vincent-van-gogh/the-starry-night-1889). All the image samples used are available in the public domain.
