Abstract
Low-light image enhancement remains a significant challenge due to the simultaneous presence of high noise, poor contrast, and color loss. This paper introduces CE-Retinex (Cross Expansion Retinex), a novel deep-learning framework that leverages a multi-dimensional cross-mixed attention mechanism and a receptive field expansion strategy. By decomposing images into reflectance and illumination components, the proposed model effectively filters noise and restores structural details, outperforming existing state-of-the-art methods in both quantitative benchmarks and visual quality.
Introduction
Capturing high-quality images in low-light environments is often hindered by a "triple threat": high sensor noise, inadequate contrast, and significant loss of color information. While traditional methods like histogram equalization are common, they frequently introduce overexposure or unnatural artifacts. Modern deep learning approaches have made strides but often suffer from "regional degradation," where certain image areas remain blurry or distorted. The authors identify a critical gap: current models often fail to filter out "negative interference" (noise and invalid background data) during enhancement, leading to the loss of fine structural details.
Methodology: The CE-Retinex Framework
The CE-Retinex algorithm is built upon the classic Retinex theory, which posits that an image is the product of an object's reflectance and its environmental illumination. The architecture is structured into three specialized modules to address these components systematically.
1. Initial Module (Decomposition)
The process begins with two separate U-shaped networks that decompose the input image into its reflectance and illuminance components. This separation allows the model to treat the intrinsic properties of the object and the lighting conditions as distinct optimization problems.
2. Optimization Module
This stage focuses on brightness adjustment and noise filtering through two primary mechanisms:
- Multi-scale Illumination Mechanism (MIM): This module adjusts brightness levels while smoothing out noise across different scales.
- Cross-mixed Attention Mechanism (CAM): As the core innovation, CAM filters information across four dimensions: channel, spatial, vertical, and horizontal. Unlike standard 2D attention, this multi-dimensional approach helps the network focus on relevant geometric features while ignoring sensor noise.
3. Detail Restoration Module
The final stage refines the image using a Receptive Field Expansion Mechanism (RFEM). By employing a multi-branch structure with dilated (or "hollow") convolutions, the model expands its perceptual field. This allows neurons to capture a wider context, aiding in the simultaneous reconstruction of large-scale structures and fine details. A specialized color consistency loss function is also applied to ensure the final output matches natural color profiles.
Experimental Results and Discussion
The model was rigorously evaluated using the LOL (Low-Light) Dataset for training and quantitative testing, alongside several testing-only datasets like SCIE, LIME, MEF, and DICM to ensure versatility across diverse conditions (e.g., nighttime aerial and indoor backlit shots).
Quantitative Performance
On the LOL dataset, CE-Retinex achieved top-tier performance metrics:
- PSNR (Peak Signal-to-Noise Ratio): 25.33
- SSIM (Structural Similarity): 0.83
- NIQE (Naturalness Image Quality Evaluator): 3.37
These results indicate that the algorithm not only preserves structural integrity but also produces images that appear natural to the human eye without requiring a "perfect" ground truth for comparison.
Qualitative Analysis
Visual comparisons demonstrate that CE-Retinex produces noticeably sharper edges and more vibrant, accurate colors compared to established models like KinD, Retinex-Net, and Zero-DCE, which often yield "washed out" or overly smoothed results.
Conclusion
CE-Retinex represents a significant advancement in low-light restoration by integrating 4D attention and expanded receptive fields. While the current model relies on paired datasets, future work will explore network self-supervision to allow the AI to learn from unpaired, real-world images. Additionally, optimizing the computational complexity of the multi-branch structures will be essential for real-time mobile applications.
Summary updated