An 11-bit High Dynamic Range (HDR) Luminance Display and Its Use to Discover Contextual Mechanisms of Real-World Luminance Normalization for Visual Acuity and Target Discrimination

Chou P. Hung, Chloe Callahan-Flintoft, Paul D. Fedele, Kim F. Fluitt, Onyekachi Odoemene, Barry D. Vaughan, Anthony J. Walker, Matthew Jaswa, Min Wei, André HarrisonbioRxiv 2019 · DOI

TL;DR

This paper introduces a novel high dynamic range (HDR) display system capable of simulating real-world lighting conditions, enabling the study of human visual perception under extreme luminance variations. They discovered a surprising form of contextual facilitation in target discrimination that depends on luminance similarity, challenging existing models of visual normalization.

What Problem Does This Paper Address?

Our visual system is remarkably adept at processing scenes with vast differences in brightness, a phenomenon known as High Dynamic Range (HDR). However, current computer vision models and research tools often struggle with these extreme luminance variations, typically operating within Standard Dynamic Range (SDR) limits (contrast ratios of 100:1 or less). This gap limits our understanding of how the human brain normalizes visual input in real-world scenarios, which is crucial for applications like autonomous target recognition and realistic display technologies. The authors aim to bridge this gap by developing and utilizing an HDR display system to investigate the underlying mechanisms of HDR luminance normalization.

Methods

The core of this research is a custom-built HDR display system. Key aspects include:

  • HDR Display Setup: A JVC DLA-RS600U 4K projector was used to achieve a luminance dynamic range exceeding 100,000:1 (650 to 0.0065 cd/m²).
  • High Bit-Depth: The system achieved pseudo 11-bit grayscale precision, offering a finer gradation of luminance levels than standard 8-bit displays.
  • Temporal Precision: A 3-ms temporal precision was maintained, crucial for dynamic visual tasks.
  • Calibration and Precision Checks: Rigorous calibration procedures were performed using a spectrophotometer to ensure accurate luminance output across the wide dynamic range. Precision checks validated the 10.7-bit per channel capability.
  • Experimental Design: Two experiments were conducted:
  • Dynamic Luminance Visual Acuity Task: A modified ETDRS LogMAR chart was used to measure visual acuity under varying background luminances and with the addition of brief, intense flashes.
  • HDR Luminance Target Discrimination Task: Participants performed a two-alternative forced-choice task to discriminate the orientation of a central Gabor target, surrounded by other Gabors under various HDR luminance conditions. This task was synchronized with EEG and eye-tracking.

Datasets

Not discussed in the paper. The experiments utilized custom-generated stimuli displayed on the developed HDR system.

Key Insights and Ideas

  • HDR Display is Crucial: Standard SDR displays are insufficient for studying real-world luminance normalization. The developed HDR system is a vital tool for this research.
  • Abrupt Darkening Induces Facilitation: A sudden decrease in luminance (e.g., from 400 to 4 cd/m²) can lead to contextual facilitation, making a central target easier to perceive, even if the target itself is high contrast. This is a novel finding, as previous research often focused on suppression in similar contexts.
  • Luminance Similarity Drives Facilitation: Surprisingly, this facilitation effect was driven by flanker stimuli that were similar in luminance to the target, not necessarily the brightest flankers. This contradicts traditional models of visual normalization (divisive or subtractive) which would predict the brightest elements to have the strongest influence.
  • Contextual Effects are HDR-Dependent: The observed facilitation effect was significantly weaker or absent under SDR conditions, highlighting the unique role of HDR in these perceptual phenomena.
  • Individual Variability in "Uniform" Conditions: When presented with a simple, uniform darkening, subjects showed a wide range of responses, from facilitation to suppression, suggesting the influence of top-down expectations or prior beliefs in visual processing.

Key Results

  • Visual Acuity: While generally decreasing with lower contrast and luminance, visual acuity was not substantially altered by brief flashes, except under the most challenging conditions (100x flash, 25% text contrast), where a mild acuity loss of up to 0.14 LogMAR was observed.
  • Target Discrimination (Behavioral):
    • In the "brightest" condition, flankers did not significantly bias choice for subject VO.
  • In the "similar" condition, flankers induced a significant facilitatory bias for subjects VO and AH, particularly with brighter adapting backgrounds (400 and 40 cd/m²).
  • The "Uniform" condition showed high individual variability, with some subjects exhibiting facilitation and others suppression.
  • Target Discrimination (EEG):
    • Peak latency of ERPs showed significant differences depending on the magnitude of luminance decrement, with larger changes leading to shorter latencies.
  • No significant flanker effect on ERP amplitude was observed, possibly due to the complexity of the stimuli and task.

Limitations and Future Work

  • Display Range: The HDR display's 100,000:1 range, while impressive, is still below the 10⁹:1 range found in some natural scenes.
  • Fixation vs. Gaze Shifts: Experiments involved maintaining fixation during luminance changes, whereas in natural scenes, these changes often occur during gaze shifts. Future work will involve controlled free-viewing to investigate this.
  • Individual Subjects: Experiments were conducted with individual subjects. Future work may explore how these findings translate to optimal teaming behavior between humans and machines.
  • Computational Modeling: The authors plan to incorporate these findings into computational models, focusing on recurrent excitation/inhibition and saliency modeling.
  • EEG Signal-to-Noise: The complexity of the visual stimuli may have introduced noise into the EEG data, potentially masking subtle effects. Future analyses might explore alternative methods.

Summary updated

Abstract

ABSTRACT Luminance can vary widely when scanning across a scene, by up to 10^9 to 1, requiring multiple normalizing mechanisms spanning from the retina to cortex to support visual acuity and recognition. Vision models based on standard dynamic range luminance contrast ratios below 100 to 1 have limited ability to generalize to real-world scenes with contrast ratios over 10,000 to 1 (high dynamic range [HDR]). Understanding and modeling brain mechanisms of HDR luminance normalization is thus important for military applications, including automatic target recognition, display tone mapping, and camouflage. Yet, computer display of HDR stimuli was until recently unavailable or impractical for research. Here we describe procedures for setup, calibration, and precision check of an HDR display system with over 100,000 to 1 luminance dynamic range (650–0.0065 cd/m^2), pseudo 11-bit grayscale precision, and 3-ms temporal precision in the MATLAB/Psychtoolbox software environment. The setup is synchronized with electroencephalography and IR eye-tracking measurements. We report measures of HDR visual acuity and the discovery of a novel phenomenon—that abrupt darkening (from 400 to 4 cd/m^2) engages contextual facilitation, distorting the perceived orientation of a high-contrast central target. Surprisingly, the facilitation effect depended on luminance similarity, contradicting both classic divisive and subtractive models of contextual normalization.
BibTeX
@misc{hung2019bit,
  author = {Chou P. Hung and Chloe Callahan-Flintoft and Paul D. Fedele and Kim F. Fluitt and Onyekachi Odoemene and Barry D. Vaughan and Anthony J. Walker and Matthew Jaswa and Min Wei and André Harrison},
  title = {{An 11-bit High Dynamic Range (HDR) Luminance Display and Its Use to Discover Contextual Mechanisms of Real-World Luminance Normalization for Visual Acuity and Target Discrimination}},
  year = {2019},
  doi = {10.1101/718437},
  howpublished = {bioRxiv},
}
Figure