We are asking whether “good composition” or “visual rightness” of visual artworks manifest themselves in particular statistical image properties. Previously, we have shown that images of artworks of Western provenance are characterized by a scale-invariant Fourier spectrum, a property which they share with natural scenes. Moreover, images of artworks are self-similar also with respect to oriented luminance gradients. Recently, we analysed the layout of edge orientations in images of >1600 oil paintings. Results showed that Shannon entropy of edge orientations was high and collinearity was low for paintings, in contrast to images of other man-made stimuli. This result suggests that large subsets of paintings are characterized by a specific layout of edge orientations. In our current work, we are studying how features learned by Convolutional Neural Networks (CNNs) in a supervised manner can be used to define a novel measure of self-similarity. Compared to the previously used method of measuring self-similarity based on oriented luminance gradients, our approach has two advantages. Firstly, we fully take color into account. Secondly, by using higher-layer CNN features, we are able to define a measure of self-similarity that relies more on image content than on basic local image features, such as luminance gradients. We are currently validating this approach by applying it to different categories of images, including artworks.
Christoph Redies, Anselm Brachmann
Experimental Aesthetics Group, Institute of Anatomy, Friedrich Schiller University Jena
Ort: Seminarraum INB
