Unsupervised detail and color restorer for Retinex-based low-light image enhancement

Retinex-based methods have demonstrated promising results in restoring low-light images to their natural, normal-light appearance. However, existing approaches often inevitably amplify hidden artifacts because the Retinex theory does not consider the various uncertain degradation patterns in dark regions. Without modeling degradations, an algorithm may easily deviate from the original color and details of regions. To address this issue, we propose a novel detail and color modeling for Retinex-based low-light image enhancement. The modeling mechanism assists our Retinex-based solution in learning rich and diverse information hidden in the dark. In addition, we develop an unsupervised loss function to reduce the solution space of Retinex decomposition. It encourages all components to mutually constrain each other, further improving the adaptiveness in unknown complex scenarios. Extensive experiments demonstrate that our approach performs favorably against state-of-the-art methods. On the SICE dataset, our method achieves 19.71 Peak Signal-to-Noise Ratio (PSNR) and 0.773 Structural Similarity Index Measure (SSIM), surpassing all compared methods in PSNR and SSIM. Our framework also generalizes robustly to the LSRW-Huawei and LSRW-Nikon benchmarks, outperforming unsupervised approaches while maintaining competitive results against supervised counterparts. The code can be accessed via: https://github.com/starsky68/DCRetinex.