Abstract Understanding the historical transformation of artistic styles implies the recognition of different stylistic properties. From a computer vision perspective, stylistic properties represent complex image features. In our work we explore the use of convolutional neural networks for learning features that are relevant for understanding properties of artistic styles. We focus on stylistic properties described by Heinrich Wolfflin in his book Principles of Art History (1915). Wolfflin identified five key visual principles, each defined by two contrasting concepts. We refer to each principle as one high-level image feature that measures how much each of the contrasting concepts is present in an image. We introduce convolutional neural network regression models trained to predict values of the five Wolfflin’s features. We provide quantitative and qualitative evaluations of those predictions, as well as analyze how the predicted values relate to different styles and artists. The outcome of our analysis suggests that the models learn to discriminate meaningful features that correspond to the visual characteristics of concepts described by Wolfflin. This indicates that the presented approach can be used to enable new ways of exploring fine art collections based on image features relevant and well-known within art history.
Paper
Full text
Learning the Principles of Art History with convolutional neural networks
Semantic Scholar · Computer Science · 2020
Abstract
Abstract Understanding the historical transformation of artistic styles implies the recognition of different stylistic properties. From a computer vision perspective, stylistic properties represent complex image features. In our work we explore the use of convolutional neural networks for learning features that are relevant for understanding properties of artistic styles. We focus on stylistic properties described by Heinrich Wolfflin in his book Principles of Art History (1915). Wolfflin identified five key visual principles, each defined by two contrasting concepts. We refer to each principle as one high-level image feature that measures how much each of the contrasting concepts is present in an image. We introduce convolutional neural network regression models trained to predict values of the five Wolfflin’s features. We provide quantitative and qualitative evaluations of those predictions, as well as analyze how the predicted values relate to different styles and artists. The outcome of our analysis suggests that the models learn to discriminate meaningful features that correspond to the visual characteristics of concepts described by Wolfflin. This indicates that the presented approach can be used to enable new ways of exploring fine art collections based on image features relevant and well-known within art history.