Learning Multi-Scale Photo Exposure Correction [Eng]

Afifi et al. / Learning Multi-Scale Photo Exposure Correction / CVPR 2021

1. Problem definition

If you have ever been interested in photography, you might be familiar with color and brightness problems, also known as exposure errors. Wrongly exposed photographs significantly affect the contrast of an image and remain the major source of issues in camera-based imaging. Generally, exposure problems can be categorized into:

• Underexposed - image is dark, so the exposure time was short

• Overexposed - image is bright with washed-out regions, so the exposure time was long

Figure 1: Underexposed image

Figure 2: Overexposed image

The main factors that impact on image exposure are: (i) shutter speed, (ii) $f$-number, the ratio of the focal length of a camera lens to the diameter of the aperture, and (iii) ISO value that controls amplification factor of received pixel signals. Once a photograph is captured and the final 8-bit standard RGB (sRGB) image is rendered, it is hard to correct exposure errors due to highly nonlinear operations applied early in the capture process by the camera image signal processor (ISP). Correcting images with exposure errors is a challenging task, even for many photo enhancement software applications.

Therefore, the exposure correction problem has been formulated as two main sub-problems: (i) color enhancement, and (ii) detail enhancement.

2. Motivation

Related work

Traditional exposure correction and contrast enhancement methods are based on adjusting image intensity values from image histograms [3,4,5,6,7]. Meanwhile, the majority of prior works adopt Retinex theory [12], which in turn formulates ineptly exposed pictures as pixel-wise multiplication of target images with correctly captured exposure settings using illumination maps. However, most of these prior methods are restricted to correcting underexpose errors only[13,14], and some cannot perform well on overexposed images.

For training deep neural networks within image enhancement tasks, it is crucially important to have large and valid paired datasets that sometimes might not be publicly available. Those that are available, focus solely on low-light underexposed images, such as Wang et. al.'s dataset[13] and the low-light (LOL) paired dataset[15].

Idea

Since most of the previous works are mainly focused on underexposed error correction and image quality enhancement, the proposed approach is the first deep learning method that explicitly correct both overexposed and underexposed sRGB images within a single model. Particularly, a coarse-to-fine deep neural network (DNN) model has been introduced. The network is fully differentiable that makes it trainable in an end-to-end manner. It starts with global color information correction, and subsequently refines image details. This approach neither reconstructs HDR images nor enhances general quality of image, it is only focused on explicitly addressing exposure errors.

Along with the coarse-to-fine DNN model, a new dataset with over 24,000 images has been introduced. The dataset images have been rendered from raw-RGB to sRGB using various exposure value (EV) settings, and broader exposure ranges compared to all previously available datasets.

Figure 3: Dataset overview.

In figure 3, from t-SNE visualization[16] it is clearly seen that the LOL[15] dataset, which is outlined in red, has relatively small number of images and covers considerably small fraction of the possible exposure levels compared to the newly proposed dataset. New dataset being based on the MIT-Adobe FiveK[17], it has generated images by adjusting the high tonal values from raw-RGB images, with a goal of realistic exposure errors emulation.

3. Method

The proposed approach is divided into two parts. First, data generation process covers all details to produce a new dataset that will be used further for training, validation and testing phases. The second part describes the proposed method of fully differentiable a coarse-to-fine deep neural network model.

Data generation

The MIT-Adobe FiveK[17] dataset contains 5,000 raw-RGB images and corresponding sRGB images that have been rendered manually by five expert photographers.

In order to emulate real exposure errors, each raw-RGB image is rendered with different exposure values (EVs). Generally, five various EV settings are used, namely $-1.5$, $-1$, $+0$, $+1$ and $+1.5$ to mimic underexposure errors, a zero-gain of original EV, and overexposure errors respectively. Zero-gain EV means original exposure settings are preserved from camera's capture time. However, note that zero-gain relative EV will not be used as a ground truth image. Specifically, images that have been rendered with corrected exposure settings by an expert photographer out of five experts[17], are being used as a target of correctly exposed images. This expert photographer is referred as an Expert C in MIT-Adobe FiveK[17].

In total, the dataset contains 24,330 8-bit sRGB images with various digital exposure settings. It is divided into three sets:

• training set of 17,675 images

• validation set of 750 images

• testing set of 5,905 images

None of the sets shares any scene in common. Images that had misalignment with their corresponding ground truth images had been excluded from the final dataset. The Figure 3 (right) illustrates generated images that mimic real exposure errors, and its corresponding properly exposed reference image.

Proposed method

Assume $I$ is a given 8-bit sRGB input image with incorrect exposure settings. The goal is to produce output image $Y$ with fewer exposure errors and noise artifacts. Since the model is expected to correct both over- and underexposed errors simultaneously, a sequential color and detail errors correction of $I$ is proposed. Particularly, rather than processing the original input image $I$, the model will be focused on a multi-resolution representation of $I$. Hence, for multi-resolution decomposition Laplacian pyramid[18] is used, which is derived from the Gaussian pyramid[19].

Let assume that $X$ is the Laplacian pyramid of image $I$ with $n$ levels, such that the last level captures low-frequency information, while the first layer depicts the high-frequency information. In this manner, one can categorize frequency levels as follows: (i) low-frequency level contains global color information; (ii) mid- and high-frequency levels stores image coarse-to-fine details.

Figure 4: Coarse-to-fine approach motivation.

The Figure 4 illustrates the general idea of coarse-to-fine approach using Laplacian pyramid. It is clearly seen that a considerable exposure correction can be achieved solely by the low-level frequency layer (Fig. 4C) of the target image (Fig. 4B). In this manner, we can enhance an entire image sequentially by correcting each Laplacian pyramid layer (Fig. 4D). However, at the inference phase, we do not have a ground truth target image for swapping the Laplacian pyramid layers. Therefore, the main goal is to predict the missing color and detail information of each Laplacian pyramid layer. Referring to these assumptions, a new coarse-to-fine deep neural network is proposed to correct global color and detail information of input image $I$ using Laplacian pyramid decomposition in a sequential manner.

Assuming that the Laplacian pyramid $X$ of image $I$ is decomposed into $n$ levels, the proposed neural network model consists of $n$ subnetworks, each being based on a U-Net-like architecture [20]. The network capacity is allocated in the form of weights depending on significance of each sub-problem that might affect the final result. The Figure 5 demonstrates the overall network model architecture.

Figure 5: Network overview architecture.

From the figure 5, we can notice that the largest U-net subnetwork (yellow), is used to process the global color information of input image $I$, in other words the last $n$th Laplacian pyramid layer. This subnetwork processes the low-frequency layer and generates an upscaled image $Y_{(n)}$. The upscaling process is done by $2 \times 2$ transposed convolutional layer with stride 2. Next, the mid-frequency layer $X_{(n-1)}$ is added to $Y_{(n)}$ and fed into the second subnetwork. This network, in turn, enhances the detail information on the current layer and produces a residual layer which is added to $Y_{(n)} + X_{(n-1)}$ with a goal of $Y_{(n-1)}$ image layer reconstruction. Generally, this entire process of refinement-upsampling is being performed until the final output image $Y$ is produced. Since the model is fully differentiable, it can be trained in an end-to-end manner.

Loss Function.

End-to-end model training has been achieved by minimizing the following loss function: $L = L_{rec} + L_{pyr} + L_{adv}$ where $L_{rec}$ denotes reconstruction loss, $L_{pyr}$ the pyramid loss, and $L_{adv}$ the adversarial loss.

Reconstruction loss. $L_1$ loss function is used between reconstructed and target reference exposed images. $L_{rec} = \sum_{p=1}^{3hw} |Y(p) - T(p)|$ $h$ and $w$ denote height and width of image, respectively, and $p$ is the index of each pixel in corrected image.

Pyramid loss. By using a simple interpolation process for upsampling operation[21], the pyramid loss is computed as follows: $L_{pyr} = \sum_{l=2}^{n} {2^{(l-2)}} \sum_{p=1}^{3 h_l w_l} |Y_{l}(p) - T_{l}(p)|$ $h_l$ and $w_l$ denote twice the height and width of $l^{th}$ level image. The pyramid loss gives a principled interpretation of the task of each subnetworks. Comparing to the training only with reconstruction loss, the pyramid loss results in less noise artifacts.

Adversarial loss. It is used as a regularizer to enhance reconstruction output of the corrected image in terms of realism. $L_{adv} = -3hwnlog(S(D(Y)))$ where $S$ is the sigmoid function, and $D$ discriminator DNN, which is trained together with the main network model.

4. Experiment & Result

Experimental setup

It is important to note that processing high-resolution images (e.g., 16-magepixel) compared to resolutions used in training phase can affect model's robustness. Therefore, bilateral guided upsampling method[22] is used to process high-res images.

During the training, the Laplacian pyramid with four levels, $n=4$ is used, to have only four subnetwork modules. The model is trained on randomly extracted patches from training images with different dimensions. First on $128 \times 128$, next $256 \times 256$ and finally $512 \times 512$ patches. Generally, the following setup was used:

• A new proposed dataset is used (section 3)

• Adam optimizer

• Training on random patches with different dimensions

• Initially train without $L_{adv}$ to speed up convergence, then add $L_{adv}$ and fine-tune the network.

Experiment results evaluation method:

• Evaluation done on test set

  • 5,905 images rendered with various EVs

• Three standard metrics for pixel-wise accuracy evaluation

  • Peak signal-to-noise ratio (PSNR) - larger value is better

  • Structural similarity index measure (SSIM) - larger value is better

  • Perceptual index ($PI$) - smaller value is better

    • $PI = 0.5(10-Ma+NIQE)$

      • $Ma$[23] and $NIQE$[24] are no-reference image quality metrics
• Baselines:

  • Histogram equalization (HE)[4]

  • Contrast-limited adaptive histogram equalization (CLAHE)[7]

  • The weighted variational model (WVM)[25]

  • The low-light image enhancement method (LIME)[14]

  • HDR CNN[26]

  • DPED models[27]

  • Deep photo enhancer (DPE) models[28]

  • The high-quality exposure correction method (HQEC)[29]

  • RetinexNet[15]

  • Deep underexposed photo enhancer (UPE)[13]

  • Zero-reference deep curve estimation method (Zero-DCE)[30]

It is worth to mention that for pixel-wise error metrics (PSNR and SSIM), during the testing stage corrected images are not compared solely to Expert C, but rather to all five expert photographers in the MIT-Adobe FiveK[17].

Result

Quantitative results

The Figure 6 demonstrates the quantitative results obtained by each method on a new introduced test set. It is clearly seen that the new proposed method achieves the best results for overexposed images, and par with set-of-the-art methods on underexposed images. The best results are colored with green and bold. The second- and third-best results are in yellow and red, respectively.

Figure 6: Quantative evaluation on a new introduced test set.

Qualitative results

The Figure 7 and 8 demonstrates the qualitative results of correcting images with exposure errors, using test set images as inputs.

Figure 7: Qualitative results of correcting images with exposure errors. Input images from test set.

Figure 8: Qualitative comparison with Adobe Photoshop’s local adaptation HDR function, and DPE

The model also generalizes well for other input images, out of the dataset. The Figure 8 shows the qualitative results of the model on public input image taken from Flickr.

Figure 9: The input images are taken from Flickr.

However, the main problem the model can face is the insufficient semantic information of input image. The Figure 10 shows when the input image has regions with complete saturation, the model cannot constrain the color inside the face due to the lack of semantic information. Additionally, when the image is completely dark, the model starts to produce remarkable noise and artifacts on the image.

Figure 10: Failure examples.

5. Conclusion

The paper proposed a single coarse-to-fine deep learning model that is able to deal with both over- and underexposed images simultaneously. Since the network is fully differentiable, it can be trained in an end-to-end manner. The main achievement of the paper is the ability to perform good on both exposure types, setting state-of-the-art results on underexposed images and par with the best methods on underexposed images.

Additionally, the key contribution of the paper is the newly proposed dataset, that contains over 24,000 images which are rendered from raw-RGB to sRGB images with various exposure settings. This dataset might be considerably useful for future work in the sphere of exposure correction problems.

Take home message

Absence of light, is often overlooked.

Abundance of light, is disgusting.

Capture correct exposure is key to the visually pleasant photographs.

Author / Reviewer information

Author

Siyavushkhon Kholmatov

• KAIST, Kim Jaechul Graduate School of AI

• Email: [e-mail]

Reviewer

TBA.

Reference & Additional materials

1. Citation of this paper

2. Official GitHub repository

3. Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing

4. Rafael C. Gonzalez and Richard E. Woods. Digital Image Processing.

5. Chulwoo Lee, Chul Lee, and ChangSu Kim. Contrast enhancement based on layered difference representation of 2D histograms..

6. Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romeny, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Com- puter Vision, Graphics, and Image Processing.

7. Karel Zuiderveld.Contrast limited adaptive histogram equalization. In Graphics Gems IV (p.474).

8. Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement.

9. Sean Moran, Pierre Marza, Steven McDonagh, Sarah Parisot, and Gregory Slabaugh. DeepLPF: Deep local parametric filters for image enhancement.

10. Jongchan Park, Joon-Young Lee, Donggeun Yoo, and In So Kweon. Distort-and-recover: Color enhancement using deep reinforcement learning.

11. Runsheng Yu,Wenyu Liu,Yasen Zhang,Zhi Qu,Deli Zhao, and Bo Zhang. DeepExposure: Learning to expose photos with asynchronously reinforced adversarial learning.

12. Edwin H Land. The Retinex theory of color vision. Scientific American.

13. Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation.

14. Xiaojie Guo, Yu Li, and Haibin Ling. LIME: Low-light image enhancement via illumination map estimation.

15. Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep Retinex decomposition for low-light enhancement.

16. Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE.

17. Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fre ́do Durand. Learning photographic global tonal adjustment with a database of input/output image pairs.

18. Peter Burt and Edward Adelson. The Laplacian pyramid as a compact image code.

19. Peter Burt. Fast filter transform for image processing. Computer Graphics and Image Processing.

20. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmentation.

21. Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion: A simple and practical alternative to high dynamic range photography.

22. Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling.

23. Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming- Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution.

24. Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer.

25. Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation.

26. Gabriel Eilertsen,Joel Kronander,Gyorgy Denes,Rafa Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs.

27. Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. DSLR-quality photos on mobile devices with deep convolutional networks.

28. Yu-Sheng Chen,Yu-Ching Wang,Man-Hsin Kao,and Yung-Yu Chuang. Deep photo enhancer: Unpaired learning for image enhancement from photographs with GANs.

29. Qing Zhang, Ganzhao Yuan, Chunxia Xiao, Lei Zhu, and Wei-Shi Zheng. High-quality exposure correction of underexposed photos.

30. Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement.