How a Trainable HVI Color Space Turns Dark Photos into Cinematic Images

Summary Overview

The authors present a novel trainable color space called HVI (Horizontal/Vertical‑Intensity) and a dedicated low‑light image enhancement network CIDNet, addressing the long‑standing issues of color bias and brightness artifacts in existing sRGB‑based LLIE methods.

Problem Statement

Color bias and brightness artifacts: sRGB‑based LLIE methods tightly couple color and intensity, leading to noticeable color shifts and halo effects under low illumination.

Red discontinuity noise and black‑plane noise: HSV, while decoupling hue, saturation, and value, introduces severe red‑color discontinuities at hue boundaries (0°/360°) and amplifies noise in extremely dark regions.

Proposed HVI Color Space

The HVI space is built on HSV but adds two key components:

Polarized HS mapping: Converts the hue‑saturation plane into orthogonal H and V components, mathematically reducing the Euclidean distance between adjacent red hues and eliminating red discontinuity noise.

Adaptive Intensity Collapse Function: A learnable function parameterized by k that compresses the intensity of low‑light regions, suppressing black‑plane noise while preserving highlight details.

These steps are illustrated in Figure 1, showing the transformation pipeline from sRGB to HVI.

HVI‑CIDNet Architecture

CIDNet is a dual‑branch encoder‑decoder network specifically designed for the HVI space:

HV branch: Models chromaticity (H and V) to separate true texture from color noise.

I branch: Processes the intensity map using the adaptive collapse function to achieve physically constrained illumination enhancement.

Lightweight Cross‑Attention (LCA): Embedded after each down‑sampling and up‑sampling stage, it enables bidirectional information flow: HV guides I to avoid over‑enhancement in dark regions, while I provides illumination weights to HV for better denoising in shadows.

The final enhanced HV and I features are concatenated and passed through a reverse HVI transform (PHVIT) to obtain the output sRGB image.

Experimental Evaluation

Extensive quantitative experiments were conducted on ten benchmark datasets (LOLv1, LOLv2‑real, LOLv2‑synthetic, DICM, LIME, NPE, MEF, VV, Sony‑Total‑Dark, SICE) and an additional LOL‑Blur joint task.

On the LOL dataset, CIDNet achieved PSNR/SSIM/LPIPS improvements over all SOTA methods (e.g., RetinexFormer, GSAD) while using only 1.88 M parameters and 7.57 GFLOPs.

On Sony‑Total‑Dark, PSNR reached 22.90 dB, a 6.68 dB gain over the second‑best method.

Although BRISQUE scores on five non‑paired datasets (DICM, LIME, NPE, MEF, VV) were slightly lower than RetinexNet, visual comparisons (Figure 4) showed CIDNet produced more realistic textures.

Cross‑method compatibility experiments (Table 3) demonstrated that inserting the HVI transform as a pre‑processing step for other LLIE models (FourLIE, GSAD) consistently raised PSNR by 1.2–3.5 dB, with GSAD+HVI achieving the best SSIM and LPIPS.

Ablation Studies

Four ablation experiments (Tables 4, Figures 5‑6) validated the contribution of each component:

Removing the adaptive intensity collapse function degrades both quantitative metrics and visual quality.

Omitting the polarized HS mapping re‑introduces red discontinuity noise.

Excluding either HV or I branch reduces the network’s ability to handle extreme illumination variations.

The LCA module is essential for synergistic denoising and illumination recovery.

Additional Results

On the combined low‑light enhancement and de‑blurring LOL‑Blur dataset, CIDNet+HVI achieved the highest scores (Table 5, Figure 7), confirming its robustness across related tasks.

Conclusion

By introducing the trainable HVI color space and the CIDNet dual‑branch architecture, the authors effectively eliminate color bias and brightness artifacts that plague traditional sRGB‑based LLIE methods. The approach delivers superior performance across a wide range of benchmarks while remaining lightweight, establishing a new state‑of‑the‑art solution for low‑light image enhancement.