Abstract
A novel paradigm for information visualization in high dynamic range images is presented in this paper. These images, real or synthetic, have luminance with typical ranges many orders of magnitude higher than that of standard output/viewing devices, thereby requiring some processing for their visualization. In contrast with existent approaches, which compute a single image with reduced range, close in a given sense to the original data, we propose to look for a representative set of images. The goal is then to produce a minimal set of images capturing the information all over the high dynamic range data, while at the same time preserving a natural appearance for each one of the images in the set. A specific algorithm that achieves this goal is presented and tested on natural and synthetic data.
Introduction
The real world exhibits a High Dynamic Range (HDR) where the luminance ratio between the darkest shadows and the brightest highlights can exceed 100,000:1. Standard display devices, such as monitors and printers, are Low Dynamic Range (LDR), typically limited to a range of roughly 255:1. Visualizing HDR data on LDR hardware requires tone mapping—a process of compressing the luminance range.
Traditional tone mapping techniques generally fall into two categories:
- Global Mapping: Applies the same transformation to every pixel. While it preserves the overall "feel" and monotonicity of the scene, it often sacrifices fine details in extreme luminance regions.
- Local Mapping: Adjusts pixels based on their neighborhood. This enhances local contrast but frequently introduces artifacts like "halos" and can invert brightness relationships (non-monotonicity).
This paper argues that for scenes with extreme dynamic ranges, a single static image is mathematically and perceptually insufficient to represent all available information without significant loss or distortion.
Proposed Methodology: The Representative Set
Instead of squeezing all information into one "muddy" image, the authors propose generating a minimal representative set of images. Each image in the set "specializes" in a specific luminance interval, effectively acting like a photographer adjusting exposure for different parts of a scene.
Luminance Segmentation and Entropy
The core of the algorithm involves analyzing the image's luminance histogram. Using a binary tree construction, the algorithm segments the histogram into regions of interest by maximizing the entropy of the distributions. This ensures that each segment contains a significant amount of visual information.
Regional Tone Mapping
The researchers adapted Larson’s histogram adjustment algorithm, a psychophysical model, to work on these specific segments rather than the whole image. By "stretching" the available 8-bit output levels (0-255) to favor a specific luminance range, they can reveal hidden details.
To select the most informative images from the generated sequence, the authors utilize Shannon Entropy. High entropy values indicate well-contrasted images with high information density. This allows for an automated selection process that minimizes redundancy.
Key Insights
- The "Movie" Paradigm: The authors suggest that HDR visualization is better served by a short video or a gallery of shots, mimicking the human eye's natural adaptation to different light levels.
- Avoiding Artifacts: Because each image in the set uses a global monotonic map, the method avoids the gradient reversals and halos common in local tone mapping.
- The Rule of 200: Empirically, the authors found that a single LDR image can comfortably display about 200 distinct luminance clusters. If a scene contains significantly more (e.g., 600 clusters), at least three images are required for a complete visualization.
Results and Evaluation
The algorithm was tested on both natural HDR images (captured via multi-exposure photography) and synthetic scenes. In a primary test case involving a cathedral interior, the algorithm identified 419 luminance clusters and generated three distinct images.
Compared to standard global tone mapping, the proposed method revealed intricate architectural details in dark areas that were previously obscured. If a single image is strictly required, the method can automatically output the frame with the highest global entropy as the "best-fit" visualization.
Conclusion
This approach shifts the focus from finding a "perfect" single compression to providing a comprehensive visual summary of HDR data. Future work includes refining temporal transitions to allow for smooth "flicker-free" viewing of the image sets and further optimizing the segmentation process to ensure the absolute minimal number of images is used.
Summary updated