High Dynamic Range Reconstruction from Saturated Visual Signals

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2893 papers1,130 conceptsfrozen 2026-10-04

Standard digital imaging sensors are defined by a fundamental constraint: the "dynamic range wall." When a scene contains light intensities that span several orders of magnitude—such as a dark interior with a sunlit window—the sensor cannot capture the full spectrum. The result is clipping, where pixels reach their maximum capacity and saturate to pure white, or fall below the noise floor into pure black. In these regions, the original visual information is not just obscured; it is physically lost.

Research in this area focuses on bridging this gap, moving from simple image capture to sophisticated reconstruction. This involves two primary strategies: "hallucinating" missing data using generative models and utilizing bio-inspired sensors that bypass the limitations of traditional frame-based imaging.

The Challenge of Irreversible Information Loss

When a sensor saturates, the underlying texture, color, and structural details are erased. As Zheyan Jin and Zhihai Xu note in their work on lens flares, restoration models face a fundamental limit in these regions because they must rely on hallucination rather than recovery. This is particularly problematic in tasks like material recovery; Tao Wen and Beibei Wang demonstrate that extreme overexposure destroys the cues necessary to estimate surface properties (SVBRDFs), as no structural data remains to guide the model.

The problem is inherently "ill-posed." There are infinite possible textures that could have existed in a blown-out highlight. Therefore, the goal of modern High Dynamic Range (HDR) reconstruction is to find the most physically plausible version of that missing data.

Generative Hallucination and Detail Synthesis

To combat clipping in single images, researchers have turned to deep learning to "hallucinate" missing details. Nima K. Kalantari and Tsang Ing Ren pioneered the use of CNNs with masked features to reconstruct HDR images from a single Low Dynamic Range (LDR) input. By training on datasets that pair LDR and HDR versions of the same scene, these models learn the statistical likelihood of what lies "behind" a saturated pixel.

This approach has been extended to specific domains like television and mobile photography. For instance, Xiangyu Chen developed HDRTVNet, which uses a divide-and-conquer strategy to convert standard broadcast content into HDR, specifically targeting the synthesis of details in clipped highlights. Similarly, Zhuoran Zheng and Wenqi Ren introduced collaborative bilateral learning to perform this reconstruction at ultra-high-definition (4K) resolutions in real-time, proving that hallucination can be both accurate and computationally efficient.

Multi-Exposure Fusion Without Motion Artifacts

The traditional way to solve the dynamic range problem is to take multiple pictures at different exposures and fuse them. However, if anything in the scene moves, this creates "ghosting" artifacts. Haesoo Chung and Nam Ik Cho proposed a "merge-and-hallucinate" framework that bypasses explicit motion compensation. Instead of trying to perfectly align pixels (which often fails in saturated regions), their model treats alignment as a brightness adjustment problem, allowing it to reconstruct ghost-free HDR images even when large-scale motion occurs in blown-out areas.

Event-Based Vision: Bypassing the Frame

Perhaps the most radical shift in this field is the move toward event-based imaging. Unlike standard cameras that capture frames at fixed intervals, event cameras (like those surveyed by Yunfan Lu and Jingao Xu) are asynchronous. Each pixel acts independently, reporting only when it detects a change in log-intensity.

This architecture offers two massive advantages:

  1. Extreme Dynamic Range: Because they operate on a logarithmic scale, event cameras can achieve over 120dB of dynamic range, capturing detail in both the heart of a shadow and the center of a bright light source simultaneously.
  2. Temporal Resolution: They operate with microsecond-level latency, effectively eliminating motion blur.

Researchers like Carl Brander and Giovanni Cioffi have leveraged these properties to create systems that can "read in the dark," performing OCR in environments where traditional RGB sensors see only black noise. Furthermore, Ling Wang demonstrated how event data can guide diffusion models to dehaze images, using the high-speed HDR signal to restore edges that are "grayed out" by low-contrast atmospheric conditions.

Beyond Intensity: Polarization and Material Cues

When intensity data is lost to glare or saturation, other properties of light can provide a fallback. Yoshiki Fukao and Shohei Nobuhara have shown that polarimetric cues can recover surface normals even on "glossy" objects where traditional stereo matching fails due to highlights. By modeling how light polarizes upon reflection, these systems can "see" the geometry of a surface even when the brightness of that surface has saturated the sensor. This multimodal approach—fusing RGB with polarization—is also being used by Haiyang Mei to detect cars in high-glare environments like nighttime driving, where direct headlight exposure would otherwise create blind spots.

Papers in this version

Showing 60 of 893 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 399 papers
  2. Version 1 · 2026-10-04 · 893 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_sensor_saturation_9af2712d,
  title = {sensor saturation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sensor-saturation/rdr_9af2712d8b294b59818cfb1d6ed40429},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_dynamic_range_reconstruction_from_s_9af2712d_v1,
  title = {High Dynamic Range Reconstruction from Saturated Visual Signals},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-dynamic-range-reconstruction-from-saturated-visual-signals/rdr_9af2712d8b294b59818cfb1d6ed40429/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}