Technical Summary: WREN - Low Light Image Enhancement Using Retinex Theory-Based Double U-Net-Like Structures

1. Problem Statement

Low-light image enhancement (LLIE) is critical for applications such as night vision, search-and-rescue, and surveillance, where traffic accidents, disasters, and crimes often occur in dark environments. Existing LLIE methods typically suffer from three primary challenges:

1. Non-trivial Retinex Decomposition: The decomposition of an image into reflectance (intrinsic color) and illumination (lighting conditions) is non-unique and ill-posed.
2. Instability in End-to-End Learning: Jointly training illumination compensation and refinement stages often leads to unstable convergence.
3. Sensitivity to Dynamic Range: Many state-of-the-art methods fail to generalize across varying lighting conditions, often resulting in over-enhancement or under-enhancement when the dynamic range of the test scene differs from the training data.

A specific limitation identified in current two-stage approaches (decomposition followed by refinement) is over-smoothing. Because the refinement stage operates on the synthesized normal-light image (the product of reflectance and illumination), it implicitly assumes noise exists in the reflectance map. However, according to Retinex theory, the reflectance map should be noise-free. Consequently, smoothing the combined image inadvertently smooths the reflectance map, degrading texture and detail.

2. Methodology: WREN Architecture

The authors propose WREN, a neural network designed to strictly adhere to Retinex theory principles to overcome over-smoothing and dynamic range sensitivity. WREN employs a double U-Net-like structure consisting of two sub-networks:

3.1. Retinex Decomposition Network ($U_1$)

This network utilizes a shared encoder with two separate decoders to decompose the input low-light image ($I_{low}$) into:

• Reflectance Map ($R$): Represents the intrinsic surface color.
• Illumination Map ($L_{low}$): Represents the incident light intensity.

Key Design Choice: Unlike other approaches that refine the reflectance map, WREN keeps the decomposed reflectance map $R$ unchanged. This strictly follows the Retinex assumption that reflectance is invariant to lighting conditions and should be noise-free.

3.2. Illumination Compensation Network ($U_2$)

This network focuses exclusively on enhancing the illumination map ($L_{low}$) obtained from $U_1$. It employs a customized U-Net architecture with the following specific components:

• Transformer Block: Inserted between the encoder and decoder to capture non-local features and promote non-local illumination smoothness, aiding in denoising.
• Attention Blocks: Integrated to focus on relevant features.
• Trainable Guided Filters: Used in the decoder to strongly denoise low-light regions while preserving edges and textures.
• Residual Blocks: Used to maintain feature integrity.

The final enhanced image is reconstructed by recombining the unchanged reflectance map ($R$) and the enhanced illumination map ($L_{normal}$) via element-wise multiplication: $I_{recon} = R \odot L_{normal}$.

3.3. Loss Functions

WREN is trained end-to-end using a composite loss function designed for robustness:

• Decomposition Loss ($L_{decomp}$): Includes a reconstruction fidelity term and a Scale-Invariant (SI) loss. The SI loss is crucial as it is mathematically invariant to global rescaling of the illumination map, providing robustness against dynamic range changes.
• Enhancement Loss ($L_{enh}$): A combination of Multi-Scale Structural Similarity (MS-SSIM), perceptual loss (VGG-16 features), CIE $L^*a^*b^*$ color distance, and a hinge penalty to prevent values from exceeding 1.0.

3. Key Contributions

1. Double U-Net Structure: A novel architecture that separates decomposition and illumination enhancement, strictly adhering to the Retinex assumption that reflectance is noise-free and invariant.
2. Scale-Invariant Training: The introduction of a scale-invariant loss function that ensures the model remains robust across different dynamic ranges and lighting conditions, addressing a common failure point in existing LLIE methods.
3. Targeted Illumination Denoising: By isolating the illumination map for enhancement using Transformer blocks and guided filters, the method avoids the over-smoothing artifacts typically caused by refining the combined image.

4. Experimental Results

The authors evaluated WREN on five datasets: LOLv1, SICE, LIEQ, NTIRE 2024, and MIT-Adobe FiveK. The model was trained on LOLv1 and tested across all datasets to demonstrate generalization.

• Quantitative Performance: WREN achieved state-of-the-art (SOTA) results across all metrics (PSNR, SSIM, LPIPS, and MAE) compared to six baselines (ReDDiT, Retinexformer, RetinexNet, SNR-Net, KinD++, Zero-DCE, and LIME). Notably, it outperformed the recent ReDDiT method by >1 dB in PSNR and >0.1 in LPIPS.
• Qualitative Performance:
  • Dark Regions: WREN performed comparably to top methods in purely dark regions.
  • Mixed Lighting: In scenes with mixed lighting (e.g., sky and dark ground), WREN successfully enhanced low-light areas without over-exposing normal-light regions. In contrast, methods like Retinexformer caused color shifts in the sky, while others (ReDDiT, RetinexNet, etc.) resulted in overexposed images.
  • Noise Handling: Unlike methods that amplify sensor noise (e.g., RetinexNet, LIME), WREN effectively suppressed noise in dark regions without introducing annoying color changes.

5. Significance and Claims

The paper claims that WREN's significance lies in its ability to achieve robust, state-of-the-art performance across multiple datasets with varying dynamic ranges by strictly following the theoretical assumptions of Retinex theory. By decoupling the refinement of illumination from the preservation of reflectance, the method eliminates the over-smoothing artifacts that limit downstream computer vision tasks. The authors emphasize that their approach provides a stable, end-to-end solution that is less sensitive to the specific dynamic ranges of training data compared to existing deep learning-based LLIE methods.