People find some images more aesthetically appealing than others. In order to generate artificial aesthetic images, we must know what statistical characteristics determine the attractiveness of pictures and how specific features of artificial neural networks can be used to produce an attractive picture.
A recent paper on arXiv.org tries to answer these questions.
A compositional pattern-producing network is used to generate a large variety of images without mimicking real ones. The researchers investigated how the parameters of the network shape the output two-point correlations. Then, human subjects evaluated the attractiveness of generated images. The used network architecture was able to represent statistics similar to a variety of natural images. It was demonstrated that humans prefer images with large correlations at scales exceeding small textured segments.
Can we generate abstract aesthetic images without bias from natural or human selected image corpi? Are aesthetic images singled out in their correlation functions? In this paper we give answers to these and more questions. We generate images using compositional pattern-producing networks with random weights and varying architecture. We demonstrate that even with the randomly selected weights the correlation functions remain largely determined by the network architecture. In a controlled experiment, human subjects picked aesthetic images out of a large dataset of all generated images. Statistical analysis reveals that the correlation function is indeed different for aesthetic images.
Research paper: Khajehabdollahi, S., Martius, G., and Levina, A., “Assessing aesthetics of generated abstract images using correlation structure”, 2021. Link: https://arxiv.org/abs/2105.08635