Deep Tone Mapping Operator for High Dynamic Range Images.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing tone mapping operators (TMOs) struggle with diverse High Dynamic Range (HDR) content and require extensive parameter tuning.
• Efficient visualization of HDR content on Low Dynamic Range (LDR) devices necessitates adaptable and fast TMOs.

Purpose of the Study:

• To propose a fast, parameter-free, and scene-adaptable deep tone mapping operator (DeepTMO) for high-resolution, high-subjective quality HDR to LDR conversion.
• To address limitations of current TMOs in handling a wide spectrum of HDR content and specific scene challenges.

Main Methods:

• Developed a deep tone mapping operator (DeepTMO) based on a conditional generative adversarial network (cGAN).
• Explored various Generator-Discriminator architectures, scale influences, loss functions, and normalization layers, adopting a multi-scale model.
• Trained the network using targets generated by the Tone Mapping Image Quality Index (TMQI) metric to leverage unlabeled HDR data.

Main Results:

• The DeepTMO successfully adapts to vast scenic content (outdoor, indoor, human, structures) and tackles HDR challenges like contrast and brightness.
• The method preserves fine-grained details and avoids common deep-learning artifacts such as blurring, tiling, and saturation issues.
• Quantitative and qualitative evaluations demonstrate high-resolution, high-quality output images across a wide range of real-world scenes.

Conclusions:

• The proposed DeepTMO provides a versatile and efficient solution for tone mapping diverse HDR content to LDR displays.
• A subjective study confirmed the method's versatility and high perceived quality, outperforming existing TMOs in adaptability and output quality.