Abstract:
|
In recent years convolutional neural networks have enjoyed great success. Especially in
the field of object recognition great leaps forward have been made. Researchers were able
to exploit the object detection features from such networks for many useful and interesting
applications like sentiment analysis and information retrieval. Unfortunately, many times
the importance of style is not being considered adequately in these systems. This is partly
because style is a concept that is difficult to define and labeled data is scarce. Recent
developments in texture synthesis and style transfer, however, sparked new interest in the
field. In particular feature correlations from convolutional neural networks, which were
trained on object recognition, have been shown to work well on these tasks. I propose
that such techniques can help in classifying style. In the course of this thesis I setup a
experiment to show that this is indeed the case. Furthermore, I show that the performance
of the CNN and the depth of the layer from which the feature correlations are taken from
influences the classification performance. |