Classification of Images as Photographs or Paintings by Using Convolutional Neural Networks

Determining whether an image is a photograph or a painting is an unsolved problem, and it is not trivially or automatically performed by humans. In previous works, humans decided which metrics should be calculated on an image to make a prediction, achieving a maximum precision of 94.82%. In this work, we propose the use of a deep learning convolutional neural network that processes the images directly, without determining the most relevant properties of an image in advance. Different modifications of the VGG network architecture are analyzed. After training the network with 16,000 images and for 100 epochs, an AUC ROC above 0.99 is achieved in images from ImageNet and in the Kaggle Painters by Numbers competition, and 0.942 in the images used by the most recent proposal in the field.