Improving Image Quality using High Dynamic Range and Aesthetics Assessment

To cope with the increasing amount of visual content available, it is important to devise automatic processes that can sort, improve, compress or store images and videos. In this thesis, we propose two different approaches to software-based image improvement.
First, we propose a study on existing aesthetics assessment algorithms. These algorithms are based on supervised neural networks.
We have collected several datasets of images, and we have tested different models using these images.We report here the performances of such networks, as well as an idea to improve the already trained networks. Our study shows that the features needed to accurately predict the aesthetics of competitive and professional are different but can be learned simultaneously by a single network.

In a second time, we propose to work with High Dynamic Range (HDR) images. We present here a new operator to increase the dynamic range of images called HDR-LFNet, that merges the output of existing operators and therefore, consists in far fewer parameters.
Besides, we evaluate our method through objective metrics and a user study. We show that our method is on-par with the state-ofthe-art according to objective metrics, but is preferred by observers during the user study, while using less resources overall.