DPEC: Dual-Path Error Compensation Method for Enhanced Low-Light Image Clarity

TL;DR

DPEC addresses noise amplification and color distortion in low-light image enhancement by replacing Retinex's multiplicative decomposition with additive error compensation. It presents a Dual-Path framework consisting of a Vision Mamba–based Brightness Error Estimator and a DenoiseCNN, trained in two stages to separately estimate brightness error and denoise while preserving texture via per-pixel error fusion $S_{DPEC} = (L+N) + E$ and $I_{final} = I_{denoise} + E(I_{LL})$. The training is guided by the HIS-Retinex loss, which combines Histogram and Reflectance Histogram losses with SSIM, perceptual, TV, inner, and smooth losses to enforce realistic brightness distribution and color fidelity. Experiments on LOL and LSRW demonstrate superior image quality (PSNR/SSIM/LPIPS) and substantially improved inference speed and memory efficiency compared with state-of-the-art methods, validating DPEC's practical impact for real-time low-light enhancement and related vision tasks.

Abstract

For the task of low-light image enhancement, deep learning-based algorithms have demonstrated superiority and effectiveness compared to traditional methods. However, these methods, primarily based on Retinex theory, tend to overlook the noise and color distortions in input images, leading to significant noise amplification and local color distortions in enhanced results. To address these issues, we propose the Dual-Path Error Compensation (DPEC) method, designed to improve image quality under low-light conditions by preserving local texture details while restoring global image brightness without amplifying noise. DPEC incorporates precise pixel-level error estimation to capture subtle differences and an independent denoising mechanism to prevent noise amplification. We introduce the HIS-Retinex loss to guide DPEC's training, ensuring the brightness distribution of enhanced images closely aligns with real-world conditions. To balance computational speed and resource efficiency while training DPEC for a comprehensive understanding of the global context, we integrated the VMamba architecture into its backbone. Comprehensive quantitative and qualitative experimental results demonstrate that our algorithm significantly outperforms state-of-the-art methods in low-light image enhancement. The code is publicly available online at https://github.com/wangshuang233/DPEC.

Figures (18)

• Figure 2: Overview of the Dual-Path Error Compensation Framework (DPEC). The Dual-Path Error Compensation Framework (DPEC) is illustrated through four diagrams: (a) first-stage training (direct error estimation), (b) second-stage training (denoising and fusion), (c) the Brightness Error Estimator (BBE) architecture, and (d) its Multi-scale Feature Fusion (MFF) submodule. These diagrams clarify the framework’s dual-path strategy: (a) trains the BBE on raw low-light inputs to predict brightness errors, followed by (b) freezing the BBE and training the DenoiseCNN (DC) to denoise inputs, with their outputs fused via addition. The BBE (c) integrates multi-scale features through its MFF submodule (d), enabling cross-scale fusion for precise error estimation. By explicitly linking (a-d) to training phases and structural components, DPEC achieves robust low-light enhancement through synergistic error correction and noise removal.
• Figure 3: The structure of DenoiseCNN and the VSS block. The DenoiseCNN focuses on removing noise from low-light images through brightness enhancement and refined convolution processing to improve image clarity. The VSS block, integrated with SS2D, enhances the overall feature extraction capability of DPEC.
• Figure 4: Scan expanding operation and Scan merging operation in SS2D. These two operations enable the mutual transformation between feature maps and image sequences.
• Figure 5: Results on LOL-v1 (top), LOL-v2-real (middle) and LOL-v2-synthetic (bottom). Our method effectively enhances the visibility and preserves the color.
• Figure 6: Results on LSRW. Other algorithms either generate over-exposed and noisy images, or introduce black spot corruptions and unnatural artifacts. While DPEC can restore well-exposed structural contents and textures.