Aesthetics in Digital Photography

7Deep Neural Network Systems

[...] as long as analytical aesthetics keeps the nature of aesthetic experience ‘on ice’, even those questions about art that it does concern itself with must remain unanswerable.

James KIRWAN (2012)

Techniques that use deep neural networks (DNN) have received a lot of attention in the last few years, every time that researchers wished to reproduce the complex decision-making carried out by an expert. The aesthetic evaluation is related to this criterion specifically, and we have thus seen several research studies rush down this path. Deng et al. (2017c) offer a good review of the work that pioneered this process, up to the first few months of 2017, and the work by Apostolidis and Mezaris (2019) and Liu et al. (2020) complements this.

As we have already said, DNNs propose an “end-to-end” approach in which the image to be evaluated is directly input into the system. The entry layer thus has one neuron per pixel, or rather, three neurons per pixel, since it is usually color images that are processed. The first few layers (a variable number between 3 and 7) are traditionally convolutional neural networks (CNNs), which make it possible to progressively reduce the size of the processed maps, by connecting a neuron from the layer n to a small number (3 × 3, 7 × 7) of neurons in the layer n − 1. The dimension of the layers can also be reduced through pooling operations, which replace the sub-sampling of convolutional layers by other decimation functions ...