Decoupling the Constraints for Generative Retinex based Image Enhancement

Abstract---Reflectance and illumination map estimation is critical in the Retinex-based low-light image enhancement(LLIE) methods. The components estimation is highly ill-posed, requiring many hand-crafted regularization terms or constraints to narrow down the solution space. However, the coupled constraints leads to a lack of flexibility in the model. The regularization terms must be redesigned and the balance parameters that controls the strength of constraints need fine-tuned in a new condition. To this end, we propose an Implicit Neural Grid Representation (INGR) to substitute Explicit Regularization Terms (ERTs) for illumination estimation, decoupling the constraints. The illumination is parameterized via bilateral grid, and a generative encoder is proposed to re-parameterize the grid. Without requirement of ERTs, our model is image-specific and more flexible. The proposed method is compared against 12 state-of-the-arts on 5 public datasets. Extensive experiments show that our method achieves sate-of-the-arts according to the qualitative and quantitative assessments. Moreover, our model outperforms competing models in terms of computing time for high-resolution image enhancement. Code is available at: \url{https://github.com/zhaozunjin/NeurBR}

Results

The results with different gamma factors values (One can use the mouse to select a sepecific image group for a closer look ).

Runing time  on high-resolution images

Code