HDR Tone Mapping Types and Selection Guide - Global to Local Comparison

HDR and Tone Mapping Fundamentals - Why It Is Needed

HDR (High Dynamic Range) images capture the vast luminance range of the real world (from sunlight at 10^9 cd/m² to starlight at 10^-3 cd/m²). However, standard displays can only show a limited dynamic range of 0-255 (8-bit), causing most HDR content to appear as blown highlights or crushed shadows without proper processing.

The role of tone mapping: Tone Mapping Operators (TMO) compress the wide dynamic range of HDR images to fit within LDR (Low Dynamic Range) display capabilities. Simple linear scaling loses contrast, so perceptually-motivated nonlinear compression considering human visual characteristics is essential.

HDR image acquisition methods:

• Exposure bracketing: Multiple shots at different exposures merged using Debevec or Robertson methods. Most common approach
• RAW development: Pseudo-HDR from 14-bit RAW data. Dynamic range approximately 12-14 EV
• HDR sensors: Sony IMX490 and similar capture 130dB+ dynamic range in single exposure
• CG rendering: Physically-based renderers output floating-point images that are inherently HDR

TMO classification: TMOs are broadly categorized as global (same transformation applied to all pixels) and local (transformation adapts based on pixel neighborhood). Global TMOs are fast and stable but limited in contrast compression; local TMOs produce higher quality but risk halo artifacts.

Global Tone Mapping - Reinhard and Logarithmic Compression

Global tone mapping applies the same transformation function to all pixels in the image. It is computationally fast and easy to implement but cannot enhance local contrast variations.

Reinhard Global TMO (2002): Inspired by photography's zone system, this is the most widely used global TMO. It first computes the log-average luminance (key value) and maps mid-tones to 0.18 (18% gray):

L_scaled = (a / L_avg) × L_world

L_display = L_scaled / (1 + L_scaled)

Parameter a (key value, default 0.18) controls overall brightness, and L_white controls maximum luminance clipping. Larger a produces brighter results, smaller a produces darker results.

Drago Logarithmic TMO (2003): Adaptively varies the logarithm base according to luminance. Dark regions use smaller bases (weaker compression) while bright regions use larger bases (stronger compression), preserving shadow detail while effectively compressing highlights. Bias parameter b (default 0.85) controls compression strength.

Filmic TMO: S-curve mimicking film characteristics, widely adopted in the game industry. John Hable's Uncharted 2 implementation is well-known, consisting of toe, linear, and shoulder segments. The ACES (Academy Color Encoding System) filmic curve is the film industry standard.

Available in OpenCV via cv2.createTonemapReinhard() and cv2.createTonemapDrago() for straightforward implementation.

Local Tone Mapping - Reinhard Local and Durand Methods

Local tone mapping adaptively determines transformation parameters based on each pixel's neighborhood information. It preserves local contrast better than global TMOs, producing more natural and detailed results, but at higher computational cost with halo artifact risks.

Reinhard Local TMO (2002): Computes surrounding luminance at multiple scales using Gaussian filters and automatically selects the optimal scale for each pixel. Larger scales enable stronger compression but crossing edges causes halos, so edge detection limits scale selection.

Durand-Dorsey Bilateral Filter TMO (2002): Separates the image into a base layer (global luminance variation) and detail layer (local texture), compressing only the base layer while preserving detail:

• Apply bilateral filter to log-luminance image to extract base layer
• Detail layer = log-luminance - base layer
• Compress base layer contrast (linear scaling)
• Add detail layer back to compressed base for reconstruction

Bilateral filter parameters: σ_spatial = 2% of image width, σ_range = 0.4 are recommended. Too-small σ_range causes halos; too-large loses detail.

Guided Filter TMO: Uses guided filter instead of bilateral filter, achieving O(N) computation for fast processing while avoiding the gradient reversal problem of bilateral filters. The ε parameter controls separation threshold, adjusted in the 0.01-0.1 range for optimal results.

Perceptual TMOs - Mantiuk and iCAM

Perceptual tone mapping explicitly models Human Visual System (HVS) characteristics to generate perceptually optimal results. These advanced methods incorporate Contrast Sensitivity Functions (CSF) and luminance adaptation models for scientifically-grounded compression.

Mantiuk TMO (2006/2008): Optimizes to minimize perceptual contrast loss based on human contrast perception models. Decomposes HDR image contrast into frequency bands and compresses each band considering CSF-based visibility thresholds:

• Multi-scale decomposition using Laplacian pyramid
• Weight contrast at each scale by CSF
• Solve optimization problem minimizing perceptual contrast loss
• Reconstruct to generate LDR image

Mantiuk '08 incorporates saturation correction to reduce color unnaturalness after tone mapping. Available via pfstools pfstmo_mantiuk08 command.

iCAM06 (Image Color Appearance Model): Extends the CIECAM02 color appearance model to image processing, integrating luminance adaptation, chromatic adaptation, and spatial frequency adaptation. Estimates adaptation luminance for each image region and performs tone mapping accordingly, producing results closer to how humans actually perceive the scene.

Display Adaptive TMO: Optimizes tone mapping considering output device characteristics (peak luminance, contrast ratio, gamma). Applies different mappings for HDR displays (1000+ nits) versus SDR displays (300 nits), maximizing each device's display capabilities.

Deep Learning HDR Processing - Single Image HDR and Learned TMOs

Deep learning has revolutionized HDR image processing in two aspects: inverse tone mapping (estimating HDR from single LDR images) and learned tone mapping operators that produce perceptually preferred results.

Single Image HDR Estimation (Inverse Tone Mapping): Technology for estimating HDR images from standard 8-bit photographs. Recovering information in blown-out regions is the primary challenge, with deep learning achieving dramatic quality improvements:

• ExpandNet (2018): Multi-scale encoder integrating local, mid-range, and global information to estimate HDR luminance. PSNR 28.5dB (evaluated with HDR-VDP-2)
• SingleHDR (2020): Simultaneously estimates camera response function for physically accurate HDR recovery
• HDRUNet (2021): U-Net based with spatial attention mechanism, significantly improving blown-out region recovery

Learned TMOs: Automatically learn perceptually preferred tone mapping from human preference data:

• Deep Tone Mapping (2020): Trained on image pairs retouched by professional photographers for artistic tone mapping
• Adaptive TMO (2022): Recognizes scene content (landscape, portrait, night) and automatically selects optimal TMO parameters

Real-time HDR processing: Lightweight deep learning models enable real-time HDR on mobile devices and game engines. Apple's Smart HDR and Google's HDR+ integrate multi-frame information as prime examples of computational photography. Processing time is approximately 100ms for 12MP on mobile GPU.

HDR photography and tone mapping books can be found on Amazon

TMO Selection Guide by Use Case and Implementation Tips

Tone mapping selection depends on use case, quality requirements, and processing speed constraints. This section provides specific recommendations per scenario and important implementation considerations.

Recommended TMO by use case:

• Fine art photography: Mantiuk '08 or Durand bilateral. Prioritize perceptual naturalness
• Real estate/architecture: Reinhard local. Balance dark interiors with bright window exteriors
• Game engines (real-time): ACES Filmic or Reinhard global. Sub-1ms GPU processing
• Film VFX: ACES workflow + custom LUT. Color management consistency priority
• Scientific/medical imaging: Logarithmic compression or Drago. Preserve data accuracy

Implementation considerations:

• Color preservation: Apply tone mapping to luminance channel only, preserving hue and saturation. Direct RGB application causes color shifts
• Gamma correction: Apply sRGB gamma (2.2) after tone mapping
• Halo prevention: Carefully tune local TMO filter parameters to suppress edge halos
• Saturation correction: Apply saturation correction factor s = 0.5-0.8 to compensate for tone mapping desaturation

OpenCV implementation: Create Reinhard TMO with cv2.createTonemapReinhard(gamma=2.2, intensity=0, light_adapt=0.8, color_adapt=0) and apply with tonemap.process(hdr_image). Results are 0-1 floating point, multiply by 255 for 8-bit conversion. Processing time is approximately 50ms for 4K images on CPU.