Extracting and Analyzing Deep Learning Features for Discriminating Historical Art: Deep Learning Features and Art

Abstract

Art historians are interested in possible methods and visual criteria for determining the style and authorship of artworks. One approach, developed by Giovanni Morelli in the late nineteenth century, focused on abstracting, extracting and comparing details of recognizable human forms, although he never prescribed what exactly to look for. In this work, we asked what could a contemporary method like convolution networks contribute or reveal about such a humanistic method that is not fully determined, but that is also so clearly aligned with computation? Convolution networks have become very successful in object recognition because they learn general features to distinguish and classify large sets of objects. Thus, we wanted to explore what features are present in these networks that have some discriminatory power for distinguishing paintings. We input the digitized art into a large-scale convolutional network that was pre-trained for object recognition from naturalistic images. Because we do not have labels, we extracted activations from the network and ran K-means clustering. We contrasted and evaluated discriminatory power between shallow and deeper layers. We also compared predetermined features from standard computer vision techniques of edge detection. It turns out that the deep network individual feature maps are highly generic and do not easily map onto obvious authorship interpretations, but in the aggregate can have strong discriminating power that are intuitive. Although this does not directly test issues of attribution, the application can inform humanistic perspectives regarding what counts as features that make up visual elements of paintings.