Abstract
A computationally fast tone mapping operator (TMO) that can quickly adapt to a wide spectrum of high dynamic range (HDR) content is quintessential for visualization on varied low dynamic range (LDR) output devices such as movie screens or standard displays. Existing TMOs can successfully tone-map only a limited number of HDR content and require an extensive parameter tuning to yield the best subjective-quality tone-mapped output. In this paper, we address this problem by proposing a fast, parameter-free and scene-adaptable deep tone mapping operator (DeepTMO) that yields a high-resolution and high-subjective quality tone mapped output. Based on conditional generative adversarial network (cGAN), DeepTMO not only learns to adapt to vast scenic-content (e.g., outdoor, indoor, human, structures, etc.) but also tackles the HDR related scene-specific challenges such as contrast and brightness, while preserving the fine-grained details. We explore 4 possible combinations of Generator-Discriminator architectural designs to specifically address some prominent issues in HDR related deep-learning frameworks like blurring, tiling patterns and saturation artifacts. By exploring different influences of scales, loss-functions and normalization layers under a cGAN setting, we conclude with adopting a multi-scale model for our task. To further leverage on the large-scale availability of unlabeled HDR data, we train our network by generating targets using an objective HDR quality metric, namely Tone Mapping Image Quality Index (TMQI). We demonstrate results both quantitatively and qualitatively, and showcase that our DeepTMO generates high-resolution, high-quality output images over a large spectrum of real-world scenes. Finally, we evaluate the perceived quality of our results by conducting a pair-wise subjective study which confirms the versatility of our method.
Introduction
High Dynamic Range (HDR) imaging captures the full breadth of luminosity found in the real world, but most consumer displays are still limited to Low Dynamic Range (LDR). To bridge this gap, Tone Mapping Operators (TMOs) are used to compress the dynamic range while preserving detail and contrast. Traditional TMOs, however, suffer from significant flaws: they are highly sensitive to manual parameter tuning and often introduce artifacts like halos or washed-out colors. DeepTMO addresses these challenges by providing a parameter-free, scene-adaptable model capable of producing high-resolution (2K) LDR images in real-time.
Methodology
DeepTMO is built upon a conditional Generative Adversarial Network (cGAN) architecture. Unlike standard GANs, this model is conditioned on the input HDR image to generate a corresponding LDR version.
Architectural Innovations
- Multi-Scale Approach: To handle high-resolution imagery (1024x2048) without tiling patterns or blurring, the authors employ a multi-scale generator and discriminator. This allows the model to simultaneously optimize for global structure and local fine-grained details.
- Hybrid Loss Function: The model moves beyond simple pixel-wise $L_1$ loss, which tends to produce blurry results. Instead, it combines Adversarial Loss, Feature Matching (FM) Loss for structural consistency, and Perceptual (VGG) Loss to preserve semantic features and textures.
- Instance Normalization: By replacing standard Batch Normalization with Instance Normalization, the network can adapt to the specific luminance statistics of a single scene, which is critical for the diverse lighting found in HDR content.
Training and Evaluation
The model was trained on a dataset of 698 HDR images. Because no single "ground truth" exists for tone mapping, the authors used 13 classical TMOs to generate candidates and selected the best target for each image using the Tone Mapping Image Quality Index (TMQI). This approach allows DeepTMO to learn a "meta-operator" strategy, selecting the most effective mapping for any given scene type.
Results and Discussion
DeepTMO achieved a mean TMQI score of 0.88, outperforming both traditional operators and previous deep learning baselines. Qualitative analysis confirms that the hybrid loss function and multi-scale architecture significantly reduce blurring compared to standard $L_1$ approaches.
In a pairwise subjective study with 20 participants, DeepTMO was preferred across diverse categories, including landscapes, structures, and indoor scenes. The model is also highly efficient, tone-mapping a 2K image in approximately 20ms, making it viable for real-time video applications.
Conclusion
DeepTMO represents a significant step forward in automated HDR visualization. While some limitations remain—such as occasional over-saturation or minor halo artifacts around extreme light sources—the model's ability to produce high-quality, high-resolution results without manual intervention makes it a versatile tool for modern display pipelines. Future work may explore unsupervised learning to further reduce dependency on existing quality metrics.
Summary updated