PROJECT TITLE :

Deep Tone Mapping Operator for High Dynamic Range Images

ABSTRACT:

The need for a rapid tone mapping operator (TMO) capable of adapting to a wide range of high dynamic range (HDR) content on low dynamic range (LDR) output devices like cinema screens or ordinary displays is necessary. Only a limited amount of HDR content can be successfully tone-mapped by existing TMOs, and the best subjective quality tone-mapped output requires substantial parameter adjustment. DeepTMO is a fast, parameter-free, and scene-adaptable deep tone mapping operator (DeepTMO) that produces a high-resolution and high-subjective quality tone mapped output for this problem. On top of adapting to enormous scenic-content (e.g., outdoor and indoor, human, structures, etc.), DeepTMO also addresses HDR-related issues such as contrast and brightness while maintaining fine-grained details, thanks to the conditional generative adversarial network (cGAN). We investigate four alternative generator-discriminator architectural designs to address some of the most common HDR-related deep-learning difficulties, such as blurring, tiling patterns, and saturation artefacts. We arrive at a multi-scale model for our problem after investigating the effects of various scales, loss functions, and normalising layers in a cGAN context. Using Tone Mapping Image Quality Index (TMIQI), an objective HDR quality indicator, we train our network to take advantage of the widespread availability of unlabeled HDR data (TMQI). Our DeepTMO produces high-resolution, high-quality output images over a wide range of real-world scenarios, which we demonstrate both numerically and qualitatively. When we undertake a pair-wise subjective evaluation of our results, we demonstrate the method's adaptability.