A Generative Adversarial Network for Tone mapping HDR images
- 0 Collaborators
A tone mapping operator converts High Dynamic Range (HDR) images to Low Dynamic Range (LDR) images, which can be seen on LDR displays. We are proposing a novel generative adversarial network to learn a combination of these tone mapping operators. In order to get pixel level accuracy, we are using residual connections between same-sized network layers. We compare this method with some of the existing tone mapping operators and observe that our method generates images with comparably high TMQI and indeed works on many different types of images. Because of the residual connections, the network can be scaled to very high dimensional images. ...learn more
Project status: Published/In Market
Groups
Student Developers for AI,
DeepLearning
Intel Technologies
AI DevCloud / Xeon,
Intel Opt ML/DL Framework
Overview / Usage
Human Visual System (HVS) can view a large dynamic range present in real world scenes. However, High Dynamic Range (HDR) images can only be seen on HDR displays. Conventional displays only support a low dynamic range (0-255) of intensity. Over time, there has been growing demand for more natural and realistic images. Nowadays, HDR images can be captured using common cameras through multi-exposure bracketing. To incorporate HDR images into an LDR display, one needs to map the higher range to a lower range. This approximation should happen in such a way that the structural similarity of the image and the naturalness are maintained. This approximation method is called tone mapping of an HDR image. Tone-mapping operators are used to achieve this process. Normally, they are classified into two categories depending on the way they process the HDR image. These categories are the global and the local tone mapping operators. We are proposing a novel generative adversarial network to learn a combination of these tone mapping operators. In order to get pixel level accuracy, we are using residual connections between same-sized network layers. We compare this method with some of the existing tone mapping operators and observe that our method generates images with comparably high TMQI and indeed works on many different types of images. Because of the residual connections, the network can be scaled to very high dimensional images.
Methodology / Approach
The main contribution of this study is the exploration of the supervised deep learn-ing techniques in the realm of HDR imaging. Because of the problems explained inthe above paragraph, the authors pose this problem as a mapping problem where do-main and co-domains are a set of images. Hence, any suitable image to image mappingmethods of deep learning should work.
Technologies Used
Intel DevCloud, Intel ColFax Cluster, Tesnorflow, Keras, PyTorch. NVIDIA for TitanX GPU card grant.