AdaInt: Learning Adaptive Intervals for 3D Lookup Tables on Real-time Image Enhancement

Image Enhancement Using Adaptive Region- Guided Multi-Step Exposure Fusion Based on Reinforcement Learning

“…NeurOp [19] has higher saturation in the enhanced image, and the proposed method is more balanced in terms of overall image brightness and contrast. On the other hand, AdaInt [18] achieves the best visual effect among all methods.…”

“…Although DeepLPF achieved good results, its predicted differentiable filter shapes are lim- Example of the retouching results on the MIT-Adobe-5K evaluation dataset. For every two rows in order: input image, CURL [42], DeepUPE [41], DeepLPF [13], FRL [26], Exposure [10], E-GAN [4], RUAS [43], PixelRL [25], NeurOp [19], A3DLUT [16], SepLUT [17], AdaInt [18], ours, and ground truth. Zoom in to better see the details.…”

“…requires multiple RRDB [45] modules to process the extracted features, which increases the overall number of parameters and computational effort of the network. In contrast, the method using the LUT class achieves the best results in terms of time overhead, number of parameters and computational effort, because the introduction of LUTs significantly reduces the number of parameters and the neural network only needs to calculate the fusion weights [16] and sampling intervals [18] of different LUTs. Compared with these methods, the proposed method does not achieve the optimal resource consumption, which may be due to the use of an improved backbone network and the exposure fusion operation making the computational and time overhead relatively large, but the method in this paper has a greater advantage in terms of interpretability.…”

“…Canqian Yang et al [17] decomposed the color transform into two parts: color component-independent and color componentcorrelated, and used one-dimensional and three-dimensional LUTs to represent them respectively, thereby improving the efficiency and quality. They also designed a 3D LUT method [18] that adaptively learns the sampling interval, and proposed a new loss function to balance fidelity and visual quality. Compared with image enhancement using LUTs, Yili Wang et al [19] proposed a novel neural color operator (NCO) for continuous image editing, and designed a new training strategy to balance the effects of different operators.…”

“…In all the experiments, ↑ higher is better, ↓ lower is better, the best result is in bold form. 20.97 0.841 0.056 Exposure [10] 18.57 0.701 0.078 PixelRL [25] 13.60 0.630 0.227 FRL [26] 16.73 0.821 0.088 Ours 24.31 0.913 0.052 DPE [40] 22.15 0.850 0.075 E-GAN [4] 15.26 0.774 0.098 DeepUPE [41] 25.22 0.895 0.053 DeepLPF [13] 21.65 0.858 0.055 CURL [42] 24.23 0.865 0.053 RUAS [43] 17.88 0.856 0.086 NeurOp [19] 21.97 0.872 0.078 A3DLUT [16] 25.12 0.922 0.051 SepLUT [17] 25.21 0.920 0.052 AdaInt [18] 25.31 0.923 0.051…”