Evaluating the Aesthetic Quality of Light Field Images Using Deep Learning

18th October 2018

Proposed by Koustav Ghosal, Sebastian Lutz
Email: {ghosalk / lutzs} at scss.tcd.ie

In this project, the primary task will be to explore novel methods to rate the aesthetic quality of light-field images. The aesthetic properties refer to those properties which makes a photograph look pleasing such as colour harmony, correct exposure, sharp focus etc. Measuring the aesthetic quality of photographic images is an active problem in computer vision. Given an image, the task of an algorithm is to predict an aesthetic rating on a given scale. However, the problem is challenging because of the subjective and overlapping nature of aesthetic properties [4]. While there has been some research in predicting the aesthetic rating of a normal photographic image [1], there has been limited research in evaluating the aesthetic rating of light-field images. With the growing interest in the research in light-field technologies, such a metric could prove useful in future.

Recently, some deep learning based aesthetic quality metrics have been proposed which train CNNs from large aesthetic visual analysis databases and then transfer that knowledge to generic images[2,3]. But, it is not trivial to adapt these algorithms to light-field image arrays because of different factors such as structure, resolution, rendering-methods etc.

In this project the tasks will cover but not limited to

  •  Understanding the existing aesthetic evaluation metrics for standard images
  • Proposing deep-learning based strategies for light field aesthetic evaluation
  • Validating the proposed method over standard light field datasets.
  • Subjective evaluation of the metric

References

1. Marchesotti, L., Murray, N. and Perronnin, F., 2015. Discovering beautiful attributes for aesthetic image analysis. International journal of computer vision , 113 (3), pp.246-266
2. Talebi, H. and Milanfar, P., 2018. Nima: Neural image assessment. IEEE Transactions on Image Processing, 27 (8), pp.3998-4011.
3. Zhang, R., Isola, P., Efros, A.A., Shechtman, E. and Wang, O., 2018. The unreasonable effectiveness of deep features as a perceptual metric. arXiv preprint.
4. Ghosal, K., Prasad, M. and Smolic, A., A Geometry-Sensitive Approach for Photographic Style Classification.