Fig 1.
The image obtained by Retinex-Net has well-preserved color details, but there is noise; the image obtained by KinD and KinD++ has good denoising effect, but the details are blurred.
Fig 2.
The framework of the proposed model.
Fig 3.
The framework of the proposed model.
Fig 4.
Multi-branch dilation convolution module.
Fig 5.
U-Net feature learning module.
Fig 6.
Adaptive iterative learning module.
Fig 7.
Layer-by-layer denoising decomposition module.
Fig 8.
Reflection component denoising module.
Fig 9.
Subjective visualization of various methods on the LOL dataset.
Table 1.
Objective evaluation results of different algorithms on LOL datasets.
Fig 10.
Subjective visualization of various methods on the LOLv2-Real dataset.
Table 2.
Objective evaluation results of different algorithms on LOL-v2-Real datasets.
Fig 11.
Subjective visualization of various methods on the DICM dataset.
Fig 12.
Subjective visualization of various methods on the MEF dataset.
Fig 13.
Subjective visualization of various methods on the LIME dataset.
Fig 14.
Subjective visualization of various methods on the NPE dataset.
Fig 15.
Subjective visualization of various methods on the Real-world dataset.
Table 3.
Objective evaluation results of different algorithms on DICM, LIME, MEF, NPE, Real-world datasets.
Fig 16.
Subjective visualization of network module ablation experiments.
Table 4.
Objective evaluation results of network module ablation experiments.
Fig 17.
Subjective visualization of the loss ablation experiment.
Table 5.
Results of objective evaluation of loss ablation experiments.
Fig 18.
Multibranch dilation convolution modules ablation experiment subjective visual map.
Table 6.
Objective evaluation results of ablation experiments for multibranch dilation convolution module.
Fig 19.
The visual effect of our method in enhancing images with both very dark and exposed areas.