Open Access
Add to BibSonomy
Abstract
Underwater images frequently suffer from color casts and poor contrast, due to the absorption and scattering of light in water medium. To address these two degradation issues, we propose an underwater image restoration method based on feature priors inspired by underwater scene prior. Concretely, we first develop a robust model to estimate the background light according to feature priors of flatness, hue, and brightness, which can effectively relieve color distortion. Next, we compensate the red channel of color corrected image to revise the transmission map of it. Coupled with the structure-guided filter, the coarse transmission map is refined. The refined transmission map preserves the edge information while improving the contrast. Extensive experiments on diverse degradation scenes demonstrate that our method achieves superior performance against several state-of-the-art methods.
© 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Underwater environment contains rich resources. Acquiring clear underwater images is of great significance for developing these resources [1–4]. However, light’s absorption and scattering seriously affect the quality of underwater images [5,6]. As shown in Fig. 1, image formation model (IFM) defined by McGlamery [7] and Jaffe [8] shows that the light received by an underwater camera can be considered as a linear combination of a direct component, a foreground scattering component, and a background scattering component. The scattering of light usually reduces the contrast of the underwater images. In addition, underwater images often suffer from color distortion in that the wavelength of red light is longer than that of green and blue. Hence, most underwater images look bluish or greenish [9]. Figs. 2(a)–2(d) shows diverse images with color distortion, loss of detail, and contrast reduction. Figs. 2(e)–2(h) are the histograms corresponding to Figs. 2(a)–2(d). The uneven distribution of the red, green, and blue channels indicates that the four images produce different degrees of color distortion [10]. Developing enhancement methods to improve underwater image quality is extremely significant.
Fig. 2. Examples of degraded underwater images. (a)-(d) are underwater degraded images. (e)-(f) are their corresponding histograms.
Model-free enhancement methods can improve the visibility of terrestrial images. These methods include histogram equalization-based, images fusion-based, and retinex-based methods. Since the underwater imaging mechanism is not considered, these methods have no significant effect on underwater images with complex and changeable imaging environments. Model-based methods can be considered as the estimations of background light and transmission map. Most of the existing model-based methods are limited to water types and scene depth, resulting in inaccurate estimations of background light and transmission map. In recent years, some methods are attempting to apply deep learning strategies to enhance underwater images [11,12]. Although image quality can be improved, the shortcomings of relying on the huge amount of training on data and longer training time make these methods inefficient.
In this work, we propose a single underwater image restoration method based on IFM, delving into a background light estimation model and the optimizer of transmission maps estimation. The proposed method embodies excellent robustness and adaptability for different types of scenes. The contributions include:
- 1) We design a background light estimation model based on feature priors of flatness, hue, and brightness, which can adaptively select the most noticeable feature according to the input image.
- 2) We propose an optimizer of transmission map estimation coupled with the structure-guided filter, which can preserve the edge information while improving the contrast.
- 3) Our method achieves superior performance against state-of-the-art methods through extensive experiments on diverse degradation scenes.
The rest of the paper is organized as follows. In Section 2, we classify the methods of processing underwater images. Section 3 details the proposed method. Section 4 reports experimental results, and conclusions are demonstrated in Section 5.
2. Related work
Researchers have proposed various underwater image processing methods to improve image quality. These methods are divided into model-free methods, model-based methods, and deep learning-based methods.
2.1 Model-free methods
Model-free methods directly adjust pixel values to improve image quality without any physical characteristics. The histogram equalization-based method is a typical enhancement method. Pizer et al. [13] developed the adaptive histogram equalization (AHE) method to improve the local contrast by transformation function to convert each pixel. Unfortunately, the AHE can amplify noise in the homogeneous region. To enhance the performance of the AHE, Reza [14] proposed the contrast limited AHE (CLAHE), which effectively avoids amplifying noise through contrast limiting. Yet, the detail enhancement around the border is not obvious. Tang. et al. [15] presented bi-histogram equalization with modified histogram bins, which can keep more image details and improve average brightness.
Fusion strategy is also widely used in image enhancement. Ancuti et al. [16] presented images and videos enhancement method by fusion, which removes the color casts and enhances the contrast of images. Li et al. [17] proposed a hybrid enhancement strategy with color correction and underwater image dehazing. Color correction is proposed to restore the color, and underwater image dehazing is used to enhance image clarity. To improve the previous research [16], an innovative method based on color balance and fusion is proposed by Ancuti et al. [18]. This method increases the exposure of dark regions and enriches the edge information. Still, some regions of the image are over-enhanced.
Another line of model-free methods is the retinex-based technology. Fu et al. [19] developed a simple RGB color projection correction algorithm for underwater images (RBE), but tends to produce over-enhanced results. Zhang et al. [20] proposed a method via extended multi-scale retinex. Bilateral filter and trilateral filter are carried out for three channels in the CIELAB color space according to features of each channel. Zhuang et al. [21] improved the visual effect of underwater images by a bayesian retinex algorithm based on multi-order gradient priors of reflectance and illumination.
Although model-free methods can improve the quality of underwater images to a certain extent, they omit influence of light’s absorption and scattering and suspended particles on imaging. These methods cannot fundamentally eliminate the impact of haze on image quality.
2.2 Model-based methods
Model-based methods are designed to assess the degradation of the image during the light is propagating, then to apply an inverse transform of the underwater image formation model [22]. The model [7,8] is given as:
where ${I^c}(x )$ is the intensity of the underwater image at the pixel x, ${J^c}(x )$ represents the restored image, ${t^c}(x )$ is the transmission map, and ${B^c}$ indicates the background light. The accurate estimations of background light and transmission map play a significant role in model-based methods. The dark channel prior (DCP) proposed by He et al. [23] is usually used to estimate the transmission map, but easily causes the incorrect restoration for underwater images. This is because the red channel is attenuated at a much higher rate than the green or blue channel in underwater medium. To improve DCP, Drew Jr et al. [24] only used the blue-green channel to obtain the transmission map of underwater images. Since the red channel is ignored, the transmission map is inaccurate, causing color distortion in the resulting images. Galdran et al. [25] inversed the red channel to obtain the red dark channel prior (RDCP) according to the absorption characteristic of wavelength. Peng et al. [26] calculated the depth map by using three different methods and selected the appropriate depth map calculation method (IBLA) according to the specific situation of underwater image. However, the contrast enhancement in some parts of the image is not obvious due to the deviation of depth map estimation. Song et al. [27] proposed a method based on statistical model of background light and optimization of transmission map, which is only suitable for the raw images with no or little red component, due to transmission map estimation depends on the information of the red channel.It is challenging to restore underwater images with different degrees of degradations. Most of the existing model-based methods are inaccurate in estimating the background light and transmission map. Some methods omit the selective absorption of light in the water medium, which degrades the result quality.
2.3 Deep learning-based methods
Recently, deep learning technology has been applied to enhance underwater images. Fabbri et al. [28] proposed a model based on generative adversarial networks (GAN) for enhancing contrast of underwater image. A CNN architecture proposed by Sharma [29], combines image enhancement tasks with classification for the first time, and utilizes the accuracy of the classification results to measure the images quality. Li et al. [30] proposed a convolutional neural model (UWCNN). This method adopts synthesized degraded images to train the corresponding UWCNN. Still, for the input image, accurately selecting the corresponding model from several UWCNN is challenging. A novel model by combining the merits of deep learning and classical histogram equalization is designed in [31], which improves image contrast without over-enhancement. Li et al. [6] presented a deep underwater image enhancement model with the multi-color space encoder network and the medium transmission-guided decoder network.
For deep learning-based methods, the quality and quantity of training data will directly impact the performance of the networks. Besides, long training and inference time is a drawback of deep learning techniques, restricting their practical application.
In general, the existing methods cannot accurately and efficiently restore the color and details of underwater images.
3. Proposed method
This section proposes a new model-based method based on feature priors, where more accurate background light and transmission map estimations are provided. We first estimate the background light based on feature priors of flatness, hue, and brightness, then refine the coarse transmission map with the structure-guided filter. The framework of the proposed method is shown in Fig. 3.
3.1 Background light estimation
The global background light represents the scattered light intensity of each color channel in the underwater scene. The typical background light estimation methods are as follows: 1) select the brightest point in the image; 2) select the farthest point in the image; and 3) select the maximum 0.1% pixel value of the dark channel. However, the above methods are not accurate in many cases, as shown in Fig. 4. Method 1) is not correct in the situation that the foreground objects or bubbles (Fig. 4(a)) are brighter than the background. Some suspended particles (Fig. 4(b)) can interfere with valid estimation with method 2). Large white objects (Fig. 4(c)) will impact the effectiveness of method 3).
Fig. 4. Some examples of the failure of the above methods. (a) foreground objects or bubbles; (b) suspended particles; (c) large white objects.
Generally, background light intensity is obtained from the smooth area away from the camera. The area around it often has the characteristics of slow brightness change and serious color deviation. So, we propose a background light estimation model based on the attributes of flatness, hue, and brightness, which defines three background lights according to different features priors and explains how they are adaptively combined. The algorithm is shown in Algorithm 1, where ξ= ${\textrm{2}^{\textrm{ - 10}}}$.
Background light $Bpart1$ based on flatness characteristic is to use quadtree segmentation algorithm to obtain the minimum variance region. We divide the input image into four quadrants, and select the quadrant to continue iterating according to the minimum scores $\textrm{S}_1$ of each area as:
A smaller value $\textrm{S}_1$ means that the flatter the region is and the slower the local brightness changes. When the iteration stops, the quadrant corresponding to the smaller $\textrm{S}_1$ is the minimum variance region of the input image. $Bpart\textrm{1}$ is obtained by averaging the minimum variance region.
where v represents the value of i that minimizes $\textrm{S}_1(i )$, $\mathrm{\Omega}(v )$ represents the minimum variance region, $m \times n$ is the number of pixels in the minimum variance region.Background light $Bpart\textrm{2}$ based on the characteristic of hue is similar to $Bpart\textrm{1}$ with the minimum variance region being replaced by the maximum color difference region and selecting the region to continue iterating according to the maximum scores $\textrm{S}_{2}$ of each area as:
A larger value $\textrm{S}_{2}$ indicates that the attenuation of the red light in the far scene is more severe than that in the close scene, and the region selected according to the largest values $\textrm{S}_{2}$ possesses the most serious color deviation. At the same time, it avoids the interference of white highlight objects. $Bpart\textrm{2}$ is calculated by averaging the maximum color difference.
where w represents the value of i that minimizes $\textrm{S}_{2}$, $\mathrm{\Omega}(w )$ represents the maximum color difference.Background light $Bpart\textrm{3}$ based on the characteristic of brightness is to take the mean value of the top 0.1% pixels of the red dark channel.
Combining $Bpart1$, $Bpart\textrm{2}$, and $Bpart\textrm{3}$, we propose a background light estimation model with adaptive parameters:
The explanation for this combined approach is as follows. If more pixels are larger than 0.5 in the grayscale image, the original image is brighter. $Bpart\textrm{3}$ is to take the mean value of the top 0.1% pixels of the red dark channel, which corresponds to the brighter points in the original image. Therefore, $Bpart\textrm{3}$ has more weight in Eq. (10) to make background light brighter. In underwater medium, the red color channel is attenuated at a much higher rate than the green and blue. If the distance between the camera and the scene point is close, the attenuation of the red component will be small. As the distance increases, the red attenuation becomes more serious, causing the image to be blue or green. At this time, the weight of $Bpart\textrm{2}$ in Eq. (10) should be larger to choose a more accurate background light.
Fig. 5 shows the background light and the restored images based on the particular characteristic of flatness, hue, or brightness and our method. In contrast, the result of the proposed method has better visual effect. Fig. 6 exhibits histogram and UCIQE of images in Fig. 5. Compared with the histogram in Figs. 6(b)-6(d), the color distribution of the proposed method is more uniform, which indicates the color cast is successfully removed. To further prove the performance of the proposed method, UCIQE [32] is used as the evaluation metric of underwater image quality, the result is shown in Fig. 6. Our UCIQE value is the highest, which signifies our results are of the best quality.
Fig. 5. An example of background light contrast. (a) Original image. (b), (c), and (d) are obtained by separately defined lights. (e) is obtained by our method.
Fig. 6. Evaluation standards for background light results. The first line displays the histogram of images in Fig. 5. The second line shows their corresponding UCIQE.
3.2 Transmission map estimation
In the process of image dehazing, DCP [23] is usually used to estimate the transmission map, but directly using DCP [23] to estimate the transmission map of underwater images easily causes to be incorrect, because red light with longer wavelength decays faster than that of green or blue in underwater medium, and misinterpretation of transmission map will lead to incomplete or excessive dehazing. Therefore, we first use RDCP [25] to estimate the coarse transmission map, then utilize the structure-guided filter to refine the transmission map.
When light travels through water, the red light has a longer wavelength than green and blue light, the red light is absorbed faster. So compensate the red channel of underwater image according to the RDCP [25], the transmission map $t(x )$ can be calculated as:
The transmission map is different for the RGB channels. The three components of the transmission map can be estimated as:
To solve block artifacts and blurry details, we design a structure-guided filter to refine the coarse transmission map, which can effectively remove block artifacts, enhance the contrast, and rich structural information of the restored image. The detailed algorithm is described in Algorithm 2.
First, a guided filter is applied to the coarse transmission map, which is decomposed into content map and detail map as:
where $grI$ is the gray image of input image as a guide image, t denotes the coarse transmission map estimated in Eqs. (12)–(14), c represents the content image obtained by the guided filter, d indicates the detail image, r is the size of the filter window, ${\varepsilon }$ represents the regularization parameter.Next, in order to rich structural information of the image, the content image is processed by a local variance filter as:
We use the guided filter on the detail image to remove the block artifacts introduced by decomposing the coarse transmission map. The operation can be expressed as:
Finally, reconstruct the refined transmission map by a linear fusion of content image C and detail image D.
Fig. 7 shows the procedure of transmission map estimation. As shown in Fig. 7(b), there are numbers of artifacts in the coarse transmission maps, and the structural information is lost. In contrast, the transmission maps (Fig. 7(c)) optimized by the structure-guided filter effectively eliminate artifacts, enrich the detailed information. The accurate estimation of transmission maps improves the contrast and visual effect of the restored images (Fig. 7(d)).
Fig. 7. Examples of transmission map estimation based on our method. (a) Original image (b) the coarse transmission map (c) the refined transmission map (d) the restored image by our method.
3.3 Image restoration
After obtaining the transmission map and the background light, the underwater image needs to be inverted to achieve the restored image. Considering that if $t(x )$ approaches 0, the value of $J(x )$ will tends to infinity, which makes restored image overflow, resulting in image distortion. Therefore, a threshold ${t_0}$ needs to be set to limit $t(x )$. The typical value ${t_0}$ is 0.1. The recovery formula can be expressed as:
The absorption coefficient $\mathrm{\lambda }g$ and $\mathrm{\lambda }b$ are related to the type of water, directly expressing the relationship between them is difficult. We use the same method as in [25] to deal with $\mathrm{\lambda }g$ and $\mathrm{\lambda }b$. The final recovery formula is:
4. Experimental results and analysis
To validate the effectiveness of the proposed method, we perform qualitative and quantitative comparisons with different methods, including RBE [19], RGHS [33], Water-Net [34], DCP [23], GDCP [35], IBLA [26]. Underwater image quality measure (UIQM) [36], underwater color image quality evaluation (UCIQE) [32], and patch-based contrast quality index (PCQI) [38] are used as the evaluation metrics to evaluate the restored results objectively. Two data sets are tested in the experiment. One is UIEB in [34] that contains 890 raw underwater images and 60 challenging images. Another is UPRC, where the images are captured by the joint laboratory of the Dalian University of Technology and Zhangzidao Group in the underwater environment of Zhangzi island in Dalian.
4.1 Qualitative comparison
In the visual comparison, the transmission map and restoration results of several model-based methods are compared. We also evaluate the visual effect of images in the diverse types of scenes.
Fig. 8 gives examples of transmission maps estimated four model-based methods. In a stingray image with no obvious separation between the front and back background, the transmission maps estimated by the four methods are similar to the visual effect of the restored images of our method. The distance between camera and stingray estimated by GDCP [35] is far, resulting in a smaller overall transmission. On the contrary, the transmission map obtained by DCP [23] is large. IBLA [26] and the proposed method estimate are more accurate. In fish’s images, the transmission maps estimated by the DCP [23] and IBLA [26] methods can hardly distinguish the foreground and the background, whose restoration images cannot remove the overall scattering effect. GDCP [35] method incorrectly reverses the foreground and the background due to the inaccurate estimation of the background light, causing the foreground of the restored image to be bright. The transmission map estimated by the proposed method can correctly express the distance relationship in the scene, retain edge information, and enhance the details of the image, and thus restoration image has an excellent visual effect.
Fig. 8. Examples of transmission maps estimated by different methods. (a) Original images (b) DCP [23] (c) IBLA [26] (d) GDCP [35] (e) the proposed method.
As shown in Fig. 9, we present some experimental results obtained by different methods. In Figs. 7(b)–7(d) and 7(h), the haze in the raw underwater images is removed by RBE [19], RGHS [33], Water-Net [34], and the proposed method, but the contrast and details of Figs. 9(b)–9(d) are not as good as those of Fig. 9(h).
Fig. 9. Qualitative comparison of image enhancement among the proposed method and state-of-the-art methods (a) Original image (b) RBE [19] (c) RGHS [33] (d) Water-Net [34] (e) DCP [23] (f) GDCP [35] (g) IBLA [26] (h) the proposed method.
The limitation of RBE [19] is to directly modify the pixels of images, without considering the reason for degradation. The results processed by this method have over-enhanced and artifacts problems, such as fishes in Image 2 and coral reefs in Image 5 (Fig. 9(b)). RGHS [33] achieves better visual quality and less noise, which benefits from relative global histogram stretching, but it fails in the challenging scene, such as Image 9 (Fig. 10(c)). Compared with other competing methods, Water-Net [34] and the proposed method can produce good visual effects. But the contrast of Water-Net [34] is relatively low, so its result is not as natural as that of the proposed method. Fig. 9(e) shows that DCP [23] method has little effect on underwater images, because the attenuation energy is not separately compensated according to different wavelengths. The result processed by GDCP [35] method shows color distortion and poor visibility, such as Image 3 (Fig. 9(f)), the background color is affected by bubbles, resulting in color distortion. Fig. 9(g) shows that IBLA [26] is suitable for the scene where the front and back scenes are separated, it can improve the brightness of the original underwater images. Yet, due to the inaccuracy of the depth map estimation, it cannot remove the effects of scattering, such as Image 2 and Image 4.
Fig. 10. The results of qualitative comparison in challenging scenes among the proposed method and state-of-the-art methods. (a) Original image (b) RBE [19] (c) RGHS [33] (d) Water-Net [34] (e) DCP [23] (f) GDCP [35] (g) IBLA [26] (h) the proposed method.
As shown in Fig. 9(e), the proposed method can achieve better visual quality because of accurate estimations of the background light and transmission map. The background light estimation model fully considers the feature priors of flatness, hue, and brightness, which can effectively solve color distortion. At the same time, it can avoid the interference of white large area objects and suspended particles. The accuracy of background light estimation is critical, directly affecting the result of transmission map estimation. Furthermore, the proposed method can also improve the contrast enrich details, which benefits from accurate estimation of transmission maps. Our structure-guided filter plays a vital role in estimating transmission maps, which can enrich more structure information while removing artifacts.
Furthermore, we tested diverse methods with underwater images captured in challenging underwater scenes to verify the proposed method’s robustness and effectiveness. As shown in Fig. 10, for Image 1,2,3 with color deviation, the disadvantages of the results processed by RBE [19], Water-Net [34], and DCP [23] are similar to those in Fig. 9. It is evident that RGHS [33], GDCP [35], and IBLA [26] still possess low contrast and color deviation. In comparison, Water-Net [34] and the proposed method increases the contrast and visibility. For Image 7 in Fig. 10, RBE [19] enlarges noise. The noise does not influence the model-based methods’ results since those methods estimate transmission maps based on local blocks [37]. However, the noise is amplified after DCP [23]. In contrast, the proposed method can avoid enlarging noise while improving the contrast and brightness.
Moreover, Fig. 11 is selected from UPRC captured in Zhangzi island. RBE [19] can solve the color deviation by adopting an effective color correction scheme and improving image quality by a variational framework-based retinex [39]. Unfortunately, the loss of details and over-enhanced results are found in Fig. 11(b), such as shells in Image 4 and Image 5. As shown in Figs. 11(c) and 11(g), the scattering effect is not completely removed in RGHS [34] and IBLA [26]. In Fig. 11(f), the over-green color and structural blurring appear in [35]. The proposed method and Water-Net [34] successfully remove color casts and promote contrast. The comparison under different types of environments shows that the proposed method is more robust and adaptable than the compared methods, and achieves better visual quality.
Fig. 11. The results of qualitative comparison in UPRC. (a) Original image (b) RBE [19] (c) RGHS [33] (d) Water-Net [34] (e) DCP [23] (f) GDCP [35] (g) IBLA [26] (h) the proposed method.
4.2 Quantitative comparison
As we know, researchers use different metrics to evaluate their results. To objectively verify the performance of the proposed method, we use several commonly used objective evaluation metrics, including PCQI, UIQM, and UCIQE.
A higher PCQI value represents the image has better contrast. The UIQM is a linear combination of colorfulness, sharpness, and contrast. A larger UIQM value means higher image quality. UCIQE is a comprehensive metric of underwater images, a higher value means the image has better balance among the chroma, saturation, and contrast.
Fig. 12 reports the average PCQI, UIQM, and UCIQE for different methods on the testing set. The average evaluation values of underwater images in Fig. 9 are shown in Fig. 12(a). Fig. 12(b) presents values of challenging underwater images appearing in Fig. 10. Furthermore, we select 100 images from the dataset UIEB [34] to test. The results are demonstrated in Fig. 12(c). Moreover, 100 images are selected from UPRC, whose results are displayed in Fig. 12(d).
Fig. 12. Quantitative results. (a) Results of images in Fig. 9 (b) results of challenging images in Fig. 10 (c) results of 100 images from the UIEB [34] (d) results of 100 images from the UPRC.
As shown in Figs. 12(a) and 12(d), it is shown that UCIQE and PCQI of images processed by the proposed method are higher than those of other methods, which means that the proposed method can better improve the visibility of restored images and balance the chroma, saturation, and contrast. UIQM of the proposed method is lower than RBE [19], since a variational framework based on retinex used in RBE [19]. However, the restored images by RBE [18] are over-enhanced as shown in Fig. 9(b), Fig. 10(b), and Fig. 11(b). PCQI, UIQM, and UCIQE of images processed by DCP [23] are close to those of original images. The result indicates that the quality of restored images is not improved. The performance produced by DCP [23] in these methods is consistent with the visual observation of low contrast and visibility in subjective analysis. It represents that haze removal with the DCP [23] cannot be directly used for underwater images. The proposed method outperforms the other methods in terms of PCQI values, which means that our method can effectively increase contrast and improve visual effects.
As shown in Fig. 12(b), the average evaluation values of images taken in the challenging scenes still have a similar trend. Fig. 12(c) shows that three evaluation values of the proposed method have the highest values compared with those of the other methods.
Compared with state-of-the-art methods, the proposed method has shown relatively good performance in qualitative and quantitative comparisons in different scenes. Satisfyingly, our approach effectively enhances image details, color information, visibility, and has a good visual effect.
5. Conclusion
In this paper, we proposed an underwater images restoration method based on feature priors, including a background light estimation model and the optimizer of transmission maps estimation. The background light model utilizes feature priors to realize adaptive adjustment, which can effectively solve color distortion. At the same time, it can avoid the interference of white large area objects and suspended particles. Considering the attenuation of the red channel is most rapid, we compensate it. Furthermore, the proposed method can also improve the contrast enrich details, which benefits from accurate estimation of transmission maps. Our structure-guided filter plays a vital role in estimating transmission maps, which can enrich more structure information while removing artifacts. Extensive evaluations on various underwater scenes validate our method can simultaneously eliminate the color distortion and enhance contrast. Qualitative and quantitative experimental results show our approach has a better performance than state-of-the-art methods.
Funding
National Natural Science Foundation of China (61702074); Natural Science Foundation of Liaoning Province (20170520196); Fundamental Research Funds for the Central Universities (3132019205, 3132019354); CAAI-Huawei MindSpore Open Fund.
Acknowledgments
Thanks to the data set provided by the joint laboratory of the Dalian University of Technology and Zhangzidao Group. We are also extremely grateful to the anonymous reviewers for their critical comments on the manuscript.
Disclosures
The authors declare that there are no conflicts of interest related to this paper.
Data availability
Data underlying the results presented in this paper are not publicly available but may be obtained from the authors upon reasonable request.
References
1. K. Liu and Y. Liang, “Underwater image enhancement method based on adaptive attenuation-curve prior,” Op. Express 29(7), 10321–10345 (2021). [CrossRef]
2. C. Li, J. Guo, and C. Guo, “Emerging from water: Underwater image color correction based on weakly supervised color transfer,” IEEE Signal Process. Lett. 25(3), 323–327 (2018). [CrossRef]
3. J. Zhou, D. Zhang, and W. Zhang, “A multifeature fusion method for the color distortion and low contrast of underwater images,” Multimed. Tools Appl. 80(12), 17515–17541 (2021). [CrossRef]
4. C. Li, C. Guo, and C. L. Chen, “Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation,” IEEE Trans. Pattern Anal. Mach. Intell., (2021).
5. Z. Duan, Y. Yuan, J. C. Lu, J. L. Wang, Y. Li, S. Svanberg, and G. Y. Zhao, “Underwater spatially, spectrally, and temporally resolved optical monitoring of aquatic fauna,” Opt. Express 28(2), 2600–2610 (2020). [CrossRef]
6. C. Li, S. Anwar, J. Hou, R. Cong, C. Guo, and W. Ren, “Underwater image enhancement via medium transmission-guided multi-color space embedding,” IEEE Trans. on Image Process. 30, 4985–5000 (2021). [CrossRef]
7. B. McGlamery, “A computer model for underwater camera systems,” Proc. SPIE 0208, Ocean Optics VI, 208 (1980).
8. J. Jaffe, “Computer modeling and the design of optimal underwater imaging systems,” IEEE J. Oceanic Eng. 15(2), 101–111 (1990). [CrossRef]
9. J. Zhou, D. Zhang, and W. Zhang, “Classical and state-of-the-art approaches for underwater image defogging: a comprehensive survey,” Front Inform Technol Electron Eng 21(12), 1745–1769 (2020). [CrossRef]
10. S. Anwar, M. Tahir, C. Li, A. Mian, F. S. Khan, and A. W. Muzaffar, “Image colorization: a survey and dataset,” arXiv preprint arXiv:2008.10774 (2020).
11. C. Li, R. Cong, S. Kwong, J. Hou, H. Fu, G. Zhu, D. Zhang, and Q. Huang, “ASIF-Net: Attention steered interweave fusion network for RGB-D salient object detection,” IEEE T. Cybern. 51(1), 88–100 (2021). [CrossRef]
12. S. Anwar and C. Li, “Diving deeper into underwater image enhancement: A survey,” Signal Processing: Image Communication 89, 115978 (2020). [CrossRef]
13. 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,” Comput. Vis. Graph. Image Process. 39(3), 355–368 (1987). [CrossRef]
14. A. Reza, “Realization of the contrast limited adaptive histogram equalization (CLAHE) for real-time image enhancement,” J. VLSI Signal Process. Syst. Signal Image Video Technol. 38(1), 35–44 (2004). [CrossRef]
15. J. Tang and N. Isa, “Bi-histogram equalization using modified histogram bins,” Appl Soft Comput 55, 31–43 (2017). [CrossRef]
16. C. Ancuti, C. O. Ancuti, T. Haber, and P. Bekaert, “Enhancing underwater images and videos by fusion,” Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. (2012), pp.81–88.
17. C. Li, J. Guo, C. Guo, R. Cong, and J. Gong, “A hybrid method for underwater image correction,” Pattern Recognit. Lett. 94, 62–67 (2017). [CrossRef]
18. C. O. Ancuti, C. Ancuti, C. De Vleeschouwer, and P. Bekaert, “Color balance and fusion for underwater image enhancement,” IEEE Trans. on Image Process. 27(1), 379–393 (2018). [CrossRef]
19. X. Fu, P. Zhuang, Y. Huang, Y. Liao, X. Zhang, and X. Ding, “A retinex-based enhancing approach for single underwater image,” Proceedings of 2014 IEEE International Conference on Image Processing (ICIP), (2015).
20. S. Zhang, T. Wang, J. Dong, and H. Yu, “Underwater image enhancement via extended multi-scale retinex,” Neuro-computing 245, 1–9 (2017).
21. P. Zhuang, C. Li, and J. Wu, “Bayesian retinex underwater image enhancement,” Eng. Appl. Artif. Intell. 101, 104171 (2021). [CrossRef]
22. K. Amer, M. Elbouz, A. Alfalou, C. Brosseau, and J. Hajjami, “Enhancing underwater optical imaging by using a low-pass polarization filter,” Opt. Express 27(2), 621–643 (2019). [CrossRef]
23. K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE Trans. Pattern Anal. Mach. Intell 33(12), 2341–2353 (2011). [CrossRef]
24. P. Drews Jr, E. do Nascimento, F. Moraes, S. Botelho, and M. Campos, “Transmission estimation in underwater single images,” Proceedings of IEEE International Conference on Computer Vision Workshops, (2013), pp. 825–830.
25. A. Galdran, D. Pardo, A. Picón, and A. Alvarez-gila, “Automatic red-channel underwater image restoration,” J. Vis. Commun. Image Represent. 26, 132–145 (2015). [CrossRef]
26. Y. Peng and P. C. Cosman, “Underwater image restoration based on image blurriness and light absorption,” IEEE Trans. on Image Process. 26(4), 1579–1594 (2017). [CrossRef]
27. W. Song, Y. Wang, D. Huang, and A. Liotta, “Enhancement of underwater images with statistical model of background light and optimization of transmission map,” IEEE Trans. Broadcast. 66(1), 153–169 (2020). [CrossRef]
28. C. Fabbri, M. J. Islam, and J. Sattar, “Enhancing underwater imagery using generative adversarial networks,” Proceedings of the IEEE International Conference on Robotics and Automation, (2018), pp. 7159–7165.
29. V. Sharma, A. Diba, D. Neven, M. S. Brown, L. V. Gool, and R. Stiefelhagen, “Classification driven dynamic image enhancement,” Proceedings of the Computer Vision and Pattern Recognition, (2018), pp.4033–4041.
30. C. Li and S. Anwar, “Underwater scene prior inspired deep underwater image and video enhancement,” Pattern Recognit. 98(1), 107038 (2019). [CrossRef]
31. X Fu and X. Cao, “Underwater image enhancement with global–local networks and compressed-histogram equalization,” Signal Processing: Image Communication 86, 115892 (2020). [CrossRef]
32. M. Yang and A. Sowmya, “An underwater color image quality evaluation metric,” IEEE Trans. on Image Process. 24(12), 6062–6071 (2015). [CrossRef]
33. D. Huang, Y. Wang, W. Song, J. Sequeira, and S. Mavromatis, “Shallow-water image enhancement using relative global histogram stretching based on adaptive parameter acquisition,” 24th International Conference on Multimedia Modeling - MMM2018, hal-01632263, 10704 (2018).
34. C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,” IEEE Trans. on Image Process. 29, 4376–4389 (2020). [CrossRef]
35. Y. Peng, K. Cao, and P. C. Cosman, “Generalization of the dark channel prior for single image restoration,” IEEE Trans. on Image Process. 27(6), 2856–2868 (2018). [CrossRef]
36. K. Panetta, C. Gao, and S. Agaian, “Human-visual-system-inspired underwater image quality measures,” IEEE IEEE J. Ocean. Eng. 41(3), 541–551 (2016). [CrossRef]
37. C. Li, J. Guo, R. Cong, Y. Pang, and B. Wang, “Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior,” IEEE Trans. on Image Process. 25(12), 5664–5677 (2016). [CrossRef]
38. S. Wang, K. Ma, H. Yeganeh, Z. Wang, and W. Lin, “A patch-structure representation method for quality assessment of contrast changed images,” IEEE Signal Process. Lett. 22(12), 2387–2390 (2015). [CrossRef]
39. R. Liu, X. Fan, M. Zhu, M. Hou, and Z. Luo, “Real-world underwater enhancement: challenges, benchmarks, and solutions under natural light,” IEEE Trans. Circuits Syst. Video Technol. 30(12), 4861–4875 (2020). [CrossRef]
