Introduction
Images captured in low-light environments present significant challenges for computer vision systems, often suffering from noise, low contrast, and severe color shifts. Traditional approaches frequently rely on Retinex theory, which decomposes an image into its intrinsic reflectance (the object's true color) and illumination (the light source). However, standard Retinex models often struggle with channel independence—processing RGB channels separately—which leads to artifacts and color bleeding. Furthermore, the decomposition is mathematically ill-posed and often fails to reflect the biological realities of human visual perception.
QRetinex-Net addresses these limitations by redefining image decomposition through hypercomplex algebra. By treating color channels as unified quaternion vectors rather than independent scalars and operating within the wavelet domain, this framework achieves state-of-the-art results in low-light enhancement and downstream computer vision tasks, boasting near-perfect reconstruction fidelity.
Methodology
The core innovation of QRetinex-Net lies in its departure from standard scalar mathematics in favor of Quaternion algebra.
Quaternion Representation
Instead of treating Red, Green, and Blue as three separate numbers, the model represents pixels as quaternions ($q = w + xi + yj + zk$). In this representation, the scalar part ($w$) corresponds to rod response (luminance), while the vector parts ($x, y, z$) represent cone responses (color). This approach treats color as a single, multi-dimensional entity, preserving inter-channel relationships through the Hamilton product and preventing the color distortion common in other models.
Wavelet-Domain Processing and Attention
The model operates in the Haar wavelet domain, allowing for the effective separation of low-frequency illumination and high-frequency texture information. A novel Symmetric Cross Attention mechanism is employed to disentangle reflectance from illumination, ensuring that structural details do not "leak" between the two branches.
Reflectance Consistency Index (RCI)
To evaluate the stability of the decomposition, the authors introduced the Reflectance Consistency Index (RCI). This metric measures how consistently a reflectance map is maintained across varying lighting conditions, with a score of 1.0 representing perfect stability. This focus on stability ensures that downstream AI tasks, such as face detection, perform reliably regardless of scene brightness.
Experimental Setup
QRetinex-Net was rigorously evaluated across several benchmarks:
- LOLv1: 500 image pairs for reconstruction accuracy.
- LCSD: Low-light crack detection in concrete structures.
- Dark Face: Face detection under challenging night-time conditions.
- LLVIP: Infrared-visible image fusion.
Results and Discussion
QRetinex-Net demonstrates substantial quantitative and qualitative improvements over existing methods. On the LOLv1 dataset, the model achieved a PSNR of 62.515 dB and an SSIM of 0.9998, significantly outperforming previous state-of-the-art models like Diff-Retinex, which sat at roughly 46.3 dB.
The model also achieved an RCI of 0.988, indicating nearly perfect reflectance stability. These improvements translated directly to downstream performance:
- Crack Detection: A ~4% improvement in F1-score over the original CrackNex.
- Face Detection: A 1.2% boost in mAP on the Wider-Face to Dark-Face transfer task.
- Image Fusion: Outperformed RDMFuse across entropy and spatial frequency metrics, leading to sharper infrared-visible fused images.
Conclusion
QRetinex-Net represents a significant leap forward in Retinex-based image decomposition. By bridging the gap between digital processing and neuroscientific principles of color perception, it provides a robust foundation for low-light computer vision. While powerful, future work will focus on optimizing computational efficiency for mobile deployment, aligning quaternion operations with perceptual color spaces like CIE Lab, and extending the framework to handle temporal consistency in video processing.
Summary updated