You are on page 1of 13

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO.

12, DECEMBER 2017 2505

Two-Layer Gaussian Process Regression With


Example Selection for Image Dehazing
Xin Fan, Member, IEEE, Yi Wang, Xianxuan Tang, Renjie Gao, and Zhongxuan Luo

Abstract— Researchers have devoted great efforts to image The dehazing algorithms that use only a single image
dehazing with prior assumptions in the past decade. Recently have attracted great attention in the past decade. Typically,
developed example-based approaches typically lack elegant mod- these methods employ strong image priors for the recovery.
els for the hazy process and meanwhile demand synthetic
hazy images by manual selection. The priors from observa- Tan [3] assumes that the haze scenes are of high smoothness
tions, and those trained from synthetic images cannot always and consistency, except for depth boundaries, and restores
reflect true structural information of natural images in practice. color contrasts by maximizing the local contrast difference.
In this paper, we present a learning model for haze removal The results are visually compelling but may be oversatu-
by using two-layer Gaussian process regression (GPR). By rated in some local regions. Fattal [1] divides the scene
using training examples, the two-layer GPR establishes a direct
relationship from the input image to the depth-dependent trans- into foreground objects and background shading under the
mission, and learns local image priors to further improve the assumption that the transmission is independent of surface
estimation. We also provide a systematic scheme to automatically shading. The method lacks a global physical model for haze
collect suitable training pairs, which works for both simu- formation and tends to underestimate the haze concentration,
lated examples and images of natural scenes. Both qualitative especially on relatively thicker haze. He et al. [2] propose
and quantitative comparisons on real-world and synthetic hazy
images demonstrate the effectiveness of the proposed approach, the dark channel prior (DCP) that the minimum of RGB
especially for white or bright objects and heavy haze regions in channels for a local patch is close to zero without haze.
which traditional methods may fail. Therefore, the local minimum can be used to estimate the
Index Terms— Example selection, Gaussian process regres- transmission. The calculation of DCP is simple yet effective,
sion (GPR), image dehazing. which gains great success for haze removal. Many variants of
DCP exist to improve the performance from various aspects.
I. I NTRODUCTION
Tarel and Hautière [4] and Gibson et al. [5] replace the
H AZE is a common phenomenon caused by small droplets
or particles suspended in the atmosphere. The images
of outdoor scenes are often degraded by haze, showing low
minimum operator with the median. Lai et al. [6] derive
the optimal transmission map from the haze model under
two scene priors. In our previous work, we [7] propose a
visibility and contrast. These degraded images create diffi-
transmission model derived from DCP that is able to edit the
culties to both human perception and automatic computer
amount of haze in an image. However, the DCP assumption
vision systems. Researches have invested significant efforts in
does not hold on sky and bright white regions that are quite
haze removal (dehazing) to restore images with high quality
common in outdoor scenes. Significant distortions inevitably
and visibility [1], [2]. We address this issue from a learning
appear in these scenarios.
perspective that trains regressors on both imaging models and
This a priori knowledge for either transmission or images
image priors from automatically selected examples for image
is built upon the field knowledge of the designers on haze
dehazing, and show that better results can be obtained by using
observations but is not necessarily true for general images.
the proposed approach.
Recently, data-driven approaches are prevailing in low-level
Manuscript received November 19, 2015; revised May 23, 2016; accepted image restoration, including denoising and super-resolution.
July 9, 2016. Date of publication July 18, 2016; date of current version In the context of image super-resolution, Huang and Wu [8]
December 13, 2017. This work was supported by the Natural Science Founda-
tion of China under Grant 61572096, Grant 61402072, and Grant 61328206. use local linear regressions to learn the relationship between
This paper was recommended by Associate Editor P. Yin. the images of high resolution and low resolution. Similarly,
X. Fan, Y. Wang, and X. Tang are with the School of Software, Dalian He and Siu [9] apply the Gaussian process regression (GPR)
University of Technology, Dalian 116024, China, and also with the Key Lab-
oratory for Ubiquitous Network and Service Software of Liaoning Province, to find the relationships. These direct regressions for the
Dalian 116621, China. formation models can render efficient image restoration.
R. Gao is with the School of Software, Dalian University of Technology, Nevertheless, these data driven approaches are not so popular
Dalian 116024, China.
Z. Luo is with the School of Software, Dalian University of Technology, in image dehazing, partially owing to the depth-dependent
Dalian 116024, China, with the Key Laboratory for Ubiquitous Network degradation process for dehazing instead of traditional con-
and Service Software of Liaoning Province, Dalian 116621, China, and also volution for image restoration. Tang et al. [10] focus on
with the School of Mathematical Sciences, Dalian University of Technology,
Dalian 116024, China. the feature selection for example-based dehazing, apply a
This paper has supplementary downloadable material available at straightforward random forest, and reported high accuracy for
http://ieeexplore.ieee.org., provided by the author. classification tasks to learn a mapping from the feature to
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. transmission. Zhu et al. [11] employ a linear model to learn the
Digital Object Identifier 10.1109/TCSVT.2016.2592328 scene depth for transmission estimation. These works directly
1051-8215 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2506 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

Fig. 1. Overview of our method. Based on the training images selected by the two-step screening process, the two-layer GPR is trained to remove haze
from a single hazy image.

bring regression models from the machine learning commu- formation and local image priors are learned from examples
nity, neglecting the structural consistency between neighboring in a unified GPR framework. GPR not only implicitly encodes
transmissions for natural scenes. An elegant trainable model the smoothness as the priors of Markov random fields (MRFs)
for this depth-dependent degradation is still highly demanded. in [3], [15], and [16], but also explicitly bridges transmission
In addition to a model characterizing the depth-dependent and observations. These direct connects circumvent iterative
degradation, it is also crucial and sometimes labor intensive inference, such as belief propagation for MRFs, and yield effi-
to find suitable examples to train the model. Tang et al. [10] cient estimation. Through the two-step screening process, good
synthesize hazy images from haze-free images and use the training data can be selected automatically. The combination of
synthetic data as the training data for the dehazing model. the two layers of GPRs further improves the proposed method
However, it is not suitable for thick haze, since it may and achieves higher accuracy. In a summary, the contributions
boost noise as the previous work. Gibson et al. [12] apply of this paper are twofold.
patch candidates with a color-dependent transformation in the 1) We propose a two-layer GPR model to learn depth-
example-based learning. This method is designed based on dependent transmission from examples for image dehaz-
3D simulation scenes, which are quite different from real- ing. This model forms a pairwise MRF, and enables
world images. Similar to those in Tang et al.’s method [10], more accurate transmission estimation attenuating ring-
it is questionable to apply the model trained from these ing artifacts while preserving local details.
synthetic examples to real-world images. In [10]–[12], syn- 2) We provide a systematic scheme to pick suitable train-
thetic hazy images from depth maps are manually chosen for ing examples from either synthetic or real-world hazy
training. However, manually choosing appropriate examples images, rendering flexible and easygoing training for the
requires the expertise both on machine learning and low-level two-layer model.
dehazing processing. The overview of our method is shown in Fig. 1, and the
In this paper, we apply a criterion against low-level image hazy images are screened for training the GPR model. The
features to pick suitable training examples either synthetic or input of the trained GPR is the multiscale feature vector of
real world for dehazing, and train straightforward yet effective the image. The output is the corresponding transmission map
classifiers for two-step example screening. We generate the of the image. Then, the haze-free image can be recovered by
training pairs on selected images using the natural transmission the haze image formation model. The selection of the training
model [7]. Moreover, we propose a two-layer GPR model examples is described in Section III, and the two-layer GPR
trainable from the selected examples. One layer of GPR is model is described in Section IV. Section V provides the
to build the regression between the transmission from rich comprehensive experimental validations both qualitatively and
features on superpixels [13], including color [10], DCP [2], quantitatively, and Section VI concludes this paper.
and structural information [14]. The other layer learns the
priors on local structures for the predicted transmissions.
II. BACKGROUND
We combine the two transmission distributions of the two
GPR layers by their product as the prediction distribution, In this section, we describe the haze image formation model
and take the mean weighted by the prediction variances of and GPR, and provide an overview of the proposed algorithm,
the two GPR layers as the transmission estimate. The physical which all lay the background of our approach.
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2507

A. Haze Image Formation Model their corresponding observations Y N = (y1 , . . . , y N )T . The


We employ the formation model for a hazy image widely conditional probability p(y ∗ | y N ) is Gaussian with the mean
accepted in computer vision and computer graphics [1], [2] m(x∗ ) and variance σ 2 (x∗ )

I ( pi ) = J ( pi )t ( pi ) + A(1 − t ( pi )) (1) p(y ∗ |y N ) ∼ N (m(x∗ ), σ 2 (x∗ )) (7)

where I ( pi ) is the observed hazy image at a pixel pi , J ( pi ) is where


the haze-free scene, A is the atmospheric light, and t ( pi ) is
the medium transmission, representing the portion of the light m(x∗ ) = kT G−1
N yN (8)
that reaches the camera with a scalar value (0 ∼ 1). The and
relationship between transmission t ( pi ) and depth d( pi ) can
be expressed as σ 2 (x∗ ) = k(x∗ , x∗ ) + σn2 − kT G−1
N k. (9)
t ( pi ) = e−βd( pi ) (2) The N-dimensional vector k is a function of the new input x∗
as k = [k(x1, x∗ ), . . . , k(x N , x∗ )]T . Hence, this predictive
where β is the scattering coefficient of the haze.
distribution of the target variable gives a Gaussian probability
We take the modified transmission model [7] that preserves
dependent on the input variable x∗ , outputting the mean and
the aerial perspective to reformulate the transmission esti-
the variance for the prediction.
mation. The modified transmission t j ( pi ) to be estimated is
GPR allows nonlinear mappings from the input to the target,
derived as
and meanwhile applies the smoothness for neighboring inputs
t j ( pi ) = ti ( pi )1−Dvisi /Dvis j (3) via the Gaussian assumption on w. We leverage GPR to build
both nonlinear mappings from input features to the transmis-
where ti ( pi ) is the original estimated transmission. Dvisi and sion and incorporate the smoothness of local transmissions.
Dvis j are the values of the maximum visibility for the input
and modified ones, respectively. The maximum visibility is
the maximum distance from which we can distinguish the III. S CREENING T RAINING E XAMPLES FOR
objects from the background in an image. We can generate G AUSSIAN P ROCESS R EGRESSION
transmissions and corresponding images with various amounts In this section, we present the process to collect the training
of haze by tuning the ratio of Dvisi and Dvis j . examples for the two-layer GPR. Training data are always
important for regressions [19]. Unfortunately, the studies on
B. Gaussian Process Regression the selection of training examples are quite rare in the lit-
GPs map multivariate normal distributions to infinite dimen- erature, especially for our case in which no accurate ground
sionality [17], [18]. The observation y of the target variable truth for transmission is available [20]. Empirical analysis on
can be expressed as a weighted summation on the basis the previous dehazing methods built upon the transmission
function φ(x) of the input x with additive noise model (1) shows that it is difficult to obtain the relationship
between the image observation and the transmission when the
y = wT φ(x) + n . (4) images have a too high or too low concentration of haze. This
The distribution on the observation assumes that weights and analysis implies that we have to select the images with the
noise behave according to a Gaussian distribution, p(w) = appropriate concentration of haze, and an image that is suitable
for training should meet two criteria: 1) the image should have
N(0, αI) and p(n ) = N(0, σn2 ), and α and σn2 are the two
hyperparameters determining the distributions of weights w haze and 2) the haze in the image should be in a range of
and observation noise n . These two parameters can be esti- certain concentration. Accordingly, we define two quantities,
i.e., h-value and t-value, and develop a two-step screening
mated by training examples. The covariance matrix of the
marginal distribution for N observations p(y) is determined process to screen training images as follows.
by the N × N Gram matrix G N with the elements
A. Screening Images With h-Value
G(xn , xm ) = k(xn , xm ) + σn2 δ(xn , xm ) (5)
To pick out hazy images, we need to be able to detect haze
where δ(·) is the Kronecker delta function, and k(·) is a kernel in the image first. Fattal [1] proposes a haze detection method,
function that relates one input vector xn to another xm . We take which uses the hue disparity between the initial image and its
the squared exponential as the kernel function expressed as semi-inverse image to detect and estimate the hazy regions in
 
−(xn − xm )2 the image. The semi-inverse image is defined as
2
k(xn , xm ) = σ f exp (6)
2l 2 c
Isemi ( pi ) = max [I c ( pi ), 255 − I c ( pi )] (10)
pi ∈I |c∈{r,g,b}
where σ 2f is the scaling parameter, and l is the length
where I is the initial image, and pi is the pixel in I . The
parameter. The training of these hyperparameters can be found hue disparity of each pixel can be obtained by computing the
in [18]. difference between the initial image and its semi-inverse image
GPR provides the distribution of the prediction y ∗ from  
a new input x∗ , given N inputs X N = (x1 , . . . , x N ) and h( pi ) =  I h ( pi ) − Isemi
h
( p i ) (11)
2508 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

Fig. 2. Screening results of the first step. Top: haze-free images in which a Fig. 3. Screening results of the second step. Top: images with dense haze.
small portion of haze at the distance appears due to the aerial perspective [2]. Bottom: images with suitable concentration of haze.
Bottom: hazy images.

where h indicates the hue channel of the image. Hue disparity minimum value of the pixels in the patch on both sides of (1),
values are normalized by dividing it by 255, and give the we have
distribution of the haze in the image. It has been observed I dark ( pi ) J dark ( pi ) ∗ t˜( pi )
that the pixels in sky or hazy areas have high intensity values = + 1 − t˜( pi ) (15)
Ac Ac
in three color channels, which cause little change between
I and Isemi . On the other hand, the pixels in the haze-free part where Ac is the atmospheric light among three color channels,
will have high hue disparity values. Therefore, we can define and the superscript dark indicates the dark channel. As the
a threshold, and label the pixel that has a hue disparity value dark channel of a haze-free image is close to 0, J dark ≈ 0, the
below the threshold as a hazy pixel. rough transmission of the local patch is given by
To investigate the overall distribution of haze in the image I dark ( pi )
more precisely, we define two features related to the hue t˜( pi ) = 1 − . (16)
Ac
disparity, which are He and Hr . We define He as
This transmission t˜ distribution map is able to approximate
1  the amount of haze in the image.
He = h( pi ) (12)
M We develop two features, Te and Tr , upon the rough
pi ∈I
transmission to represent the overall concentration of haze in
where M is the number of pixels in I , and h( pi ) is the hue the image. We define Te as
disparity value of the pixel pi . The feature can be used to 1 
represent the mean concentration of the haze in the image. Te = t˜( pi ) (17)
M
We use Hr to represent the ratio of hazy pixels, defined as pi ∈I

1  where M is the number of the pixels in I . Te is the average


Hr = r (h( pi )) (13) transmission of the image, representing the global concentra-
M
pi ∈I tion of the haze in the image. The other feature Tr is defined as
where r (h( pi )) is defined as 1 
 Tr = r (t˜( pi )) (18)
M
1, h( pi ) ∈ [0, th ] pi ∈I
r (h( pi )) = (14)
0, otherwise where r (t ( pi )) is defined as

where th is a predefined threshold. Since the pixels that have 1, t˜( pi ) ∈ [tl , 1]
hue disparity values below the threshold th are hazy pixels, this r (t ( pi )) = (19)
0, otherwise
feature represents the ratio of the hazy pixels to the total. We
can then show the overall distribution of the haze in the image, where tl is a predefined threshold. The pixel with the rough
and apply support vector machines (SVM) [21] to separate transmission below tl has a high concentration of haze. Tr rep-
the hazy images from other images based on the two features, resents the proportion of the pixels with transmission above tl
since they are effective, and SVM performs quite well on this and provides the ratio of regions with suitable concentration of
type classification problem. Fig. 2 shows the screening results haze in the image. Again, we use SVM to classify the images
of the first step. We can see that the images can be classified with the two features as the input. As shown in Fig. 3, this
accurately and hazy images can be identified. process can accurately identify dense hazy images from the
others. Subsequently, the images with suitable concentration
(not dense) of haze are chosen as training examples.
B. Screening Images With t-Value It is difficult to choose the existing dehazing method to
When the concentration of the haze in the image is too high, generate the target transmission maps for the train images,
the image is also not suitable for training. Hence, we need to since not a single dehazing method achieves the best dehazing
exclude these images when we prepare training data. We use results for all hazy images. We found that Fattal’s method [1],
the DCP [2] to roughly estimate the transmission of the image He et al.’s method [2], and Tang et al.’s method [10] suffice
for evaluating the haze concentration. to produce target transmission maps for images with different
Assuming that the rough transmission of the pixels in a levels of haze. He et al.’s method works better for relatively
local patch is a constant, we denote it by t˜( pi ). Taking the dense hazy images than Tang et al.’s method, though color
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2509

GPR training include the hue, multiscale dark channel, local


maximum contrast, saturation, and Gabor features [10], [14].
Hue disparity can be used to detect haze [23]. In this paper,
we use the hue disparity between the hazy image and its semi-
inverse image as the feature for our training. The multiscale
dark channel, local maximum contrast, and saturation can be
expressed as
D s ( pi ) = min (min(I c ( ps ))) (20)
ps ∈s ( pi ) c∈{r,g,b}

Fig. 4. Training images. Top: hazy images. Bottom: dehazed results. (min ps ∈s ( pi ) (minc∈{r,g,b} (I c ( ps ))))2
C ( pi ) =
s
1−
distortions may appear in sky regions. Tang et al.’s dehazing (max ps ∈s ( pi ) (maxc∈{r,g,b} (I c ( ps ))))2
results have higher visual quality and fewer color distortions, (21)
especially for moderate hazy images. Fattal’s method tends to s (min ps ∈s ( pi ) (minc∈{r,g,b} (I c ( ps ))))
S ( pi ) = 1 − (22)
underestimate the haze, but it preserves the structure of the (max ps ∈s ( pi ) (maxc∈{r,g,b} (I c ( ps ))))
image well for thin hazy images. We pick out hazy images
by the proposed screening, and categorize these hazy images where D s is the dark channel, C s is the local maximum
into different levels of haze as the byproduct of screening contrast, and S s is the saturation. I c denotes the color RGB
classifiers. Subsequently, we apply these three dehazing meth- channels of the input hazy image I . s is a local window
ods respective to the haze level of an image in order to centered at the pixel pi with size s × s, where pi denotes
generate corresponding transmission map for the image. These a pixel in the image I . A small s leads to an overdehazing
haze/transmission pairs constitute the training examples for the result, and on the contrary, a large s results in a large amount of
proposed two-layer model. We list some of the hazy images haze and halo effects near depth edges. Thus, we apply a mul-
with different levels of haze and their corresponding dehazed tiscale s into dark channel features. In this paper, we use four
outputs for training, as shown in Fig. 4. scales as D̃ = [D 1 , D 4 , D 7 , D 10 ], C̃ = [C 1 , C 4 , C 7 , C 10 ],
and S̃ = [S 1 , S 4 , S 7 , S 10 ].
IV. T WO -L AYER G AUSSIAN P ROCESS Gabor filtering performs well in the field of textural repre-
R EGRESSION FOR D EHAZING sentation and classification, and textural features in an image
can be apparently changed by haze. We convolve a given
In this section, we briefly describe the oversegmentation for
image with Gabor filter and calculate the Gabor features [14]
superpixels and the multiscale feature vectors on superpixels,
from the filtered image. Then, we apply Gabor features to
generating the inputs of our model. Subsequently, we detail
extend the information content of the feature vectors. Different
the two-layer GPR model for dehazing and the subsequent
scales and orientations represent different image information.
processing of transmissions for the haze removal.
We use Gabor features with three scales and eight directions
in our experiment. More details can be found in [14].
A. Superpixel
We employ the superpixel in [13] to assemble the feature
vectors. The superpixels are obtained by using the following C. Two-Layer Gaussian Process Regression
features: We build a two-layer GPR model for learning transmissions
1) inter-region texture similarity; on superpixels. As shown in Fig. 5, the first GPR layer takes
2) intra-region texture similarity; multiscale feature vectors as the input, and find the conditional
3) inter-region brightness similarity; probability of transmission dependent on the input features
4) intra-region brightness similarity; from training examples. Center pixels can be represented by
5) inter-region contour energy; the neighboring pixels or a subset of the connected neigh-
6) intra-region contour energy; boring pixels [25]. The second GPR relates the transmission
7) curvilinear continuity. of the current superpixel with those of neighboring superpix-
This combination of features segments superpixels well. The els. These two layers form an MRF, where the conditional
pixels in a superpixel present homogenous texture, brightness, probability densities are Gaussian with the mean and variance
and color. The use of these midlevel features, other than determined by training examples. Hence, the inference of
image patches, renders a lower dimensional input space that transmissions for superpixels is quite straightforward, which
favors smaller training examples and simpler regressors, and we detail in the following.
meanwhile partially provides structural contexts. The detailed 1) Gaussian Process Regression From Feature to
discussions about superpixels and local patches for dehazing Transmission: We use the average feature vector within
can be found in [22]. the superpixel of interest as the input fs of the first GPR layer
1 
B. Multiscale Feature Vector fs = F( pi ) (23)
|n|
The ability to incorporate multiscale features is one of pi ∈Si
the advantages of regression methods. The features for our F( pi ) = [H, D̃, C̃, S̃, G f 3,8 ] (24)
2510 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

probability, and omit the training observations y N in the


probability of (7), as this term is fixed for a given training
set. We substitute the feature vector fs∗ in (23) as x∗ to (8)
and (9), and take the mean m(fs∗ ) and the variance σ 2 (fs∗ ) as
the transmission estimation and its error predicted from the
observed feature at every superpixel.
The predicted transmission of the first GPR layer is able to
roughly reflect the depth and global structures of an image,
as shown in Fig. 6(b), but it presents local disparity and the
superpixel lack of interactions with its neighbors.
2) Gaussian Process Regression on Neighboring
Transmissions: We use the second GPR layer to smooth the
local discrepancies among predicted transmissions. These
discrepancies destroy the consistency of image structures, and
may lead to unnatural estimation in local regions. This GPR
layer plays a similar role to the connections between latent
variables of the MRFs in [15] and [16], but needs no iterative
energy optimization or inference process. We can achieve the
smoothing by a one-step prediction.
The target is the transmission t (Si ) at a superpixel Si , and
the collection of the training transmissions is also the same
as that for the first layer. We train the second GP regressor
Fig. 5. Two-layer GPR model of transmissions for dehazing. The first layer by using every transmission of training images as the target y
GPR1 learns the mapping from the midlevel features to transmission, and and their nearest eight neighbors as the input vector tn
the second layer GPR2 provides the relationship of transmissions between
neighboring superpixels. tn = [t (S1 ), . . . , t (S j ), . . . , t (S8 )] S j ∈Ne (Si ) . (26)

where Si is the superpixel, n is the number of pixels in Si , and It should be noted that a superpixel does not necessarily
F( pi ) is the multiscale feature extracted at the pixel pi . The have a fixed number of positional adjacent neighbors as a
feature vector includes H , D̃, C̃, S̃, and G f 3,8 , representing pixel does. Thus, we take the eight nearest neighbors, not
the hue, DCP with four scales, contrast with four scales, necessarily sharing a boundary with (nor being adjacent to) the
saturation with four scales, and Gabor features with three current superpixel, to generate input vectors with an identical
scales and eight orientations, respectively, as described in dimension for the regressor.
Section IV-B. The recent work [10] constructs a 325D feature In the prediction stage, we concatenate the transmissions t ∗f
vector at every pixel in a 5 × 5 patch. We take the average predicted in the first layer of the nearest eight neighbors as
of the pixel feature vectors within a superpixel where these the new input tn∗ = [t ∗f (S1 ), . . . , t ∗f (S j ), . . . , t ∗f (S8 )] S j ∈Ne (Si )
vectors share similar characteristics. Finally, we have a 37D for the layer. Similarly, we obtain a Gaussian probability
feature vector for each superpixel, significantly decreasing the for the predicted transmission t ∗ dependent on that of its
regression complexity. neighbors tn∗ as
We take the transmission t at a superpixel as the target     
p t ∗ tn∗ ∼ N m tn∗ , σ 2 tn∗ . (27)
variable y for the first layer. However, it is nontrivial to collect
the target transmission values for training the GP regres- Again, the mean estimation and variance can be calcu-
sor from real-world hazy images, as these values are not lated by substituting the neighboring input transmissions
directly available from observations. We apply three dehazing to (8) and (9). As shown in Fig. 6(c), this transmission
methods [1], [2], [10] to the training images, and use the trans- prediction of a superpixel from its neighbors shows global
mission map that produces the best restoration as the target smoothness and consistency.
map. We then take the feature vector at every superpixel as 3) Transmission Estimation With the Two-Layer GPR: The
the input x and the corresponding transmission as the target y two-layer model in Fig. 5 for transmissions at superpixels
to train the parameters of the GP regressor in (5) and (6). forms an MRF similar to [15] and [16]. In the spirit of classical
We are able to have the conditional probability of the trans- belief propagation inference on MRF [15], [26], the marginal
mission prediction t ∗f , given any new input feature vector fs∗ , probability of a node in MRF, herein the transmission at a
when N training transmissions t f and their corresponding superpixel, is proportional to the product of messages reflect-
inputs are available. As shown in (7), the conditional prob- ing the conditional dependences on adjoining nodes or factors
ability is a Gaussian distribution (observed features and neighboring transmissions in Fig. 5).
     These conditional dependences are given by Gaussian proba-
p t ∗f fs∗ ∼ N m fs∗ , σ 2 fs∗ . (25) bility density functions, as shown in (25) and (27). Therefore,
we have the transmission estimation
We explicitly introduce the input vector fs∗ , on which the mean      
and the variance depend, into the notation of the conditional p(t ∗ ) ∝ N m tn∗ , σ 2 tn∗ N m fs∗ , σ 2 fs∗ . (28)
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2511

Fig. 6. Haze removal with two-layer GPR. (a) Hazy image. (b) Transmission image estimated by the first layer GPR. (c) Transmission image estimated by
the second-layer GPR. (d) Refined transmission image estimated by the combination of the two-layer GPR. (e) Final transmission image improved by guided
filtering [24] and Gaussian filtering. (f) Our final haze removal result. (g) Original haze-free image.

The distribution of the first layer N (m(fs∗ ), σ 2 (fs∗ )) indicates variance indicates that the transmission estimate from the
the conditional dependence on the observed feature of the two-layer model yields the accuracy improvements for
current superpixel, while the second one N (m(tn∗ ), σ 2 (tn∗ )) transmission estimation. The quantitative analysis is in
characterizes the interaction of the transmissions between the Section V-A.
current superpixel and its neighboring ones. Alternatively, it is possible to jointly train one GP regression
The distribution of p(t ∗ ) in (28) is a scaled Gaussian from the 333D (37 × 9) augmented feature space to the
distribution 9D transmission space (one center superpixel and its eight
neighbors), generating an R333×9 mapping. It is known that
p(t ∗ ) ∼ N (m(t ∗ ), σ 2 (t ∗ )) (29)
the training complexity of GPR is exponential to the mapping
and the mean and the variance of p(t ∗ ) are given dimensionality. The decomposition into two simpler GPR
as [27], [28] mappings (one maps 37D to 1D, and one from 8D to 1D)
    in our proposed model significantly reduces the complexity.
∗ m tn∗ σ 2 fs∗ + m fs∗ σ 2 tn∗
m(t ) =   (30) Moreover, both layers of GPR have clear interpolations:
σ 2 fs∗ + σ 2 tn∗ one relating the feature to transmission and the other
and encoding the interactions between neighboring transmissions,
  respectively.
∗ σ 2 fs∗ σ 2 tn∗
σ (t ) = 2  ∗
2  . (31) The transmission output is superpixelwise. For a further
σ fs + σ 2 tn∗ refinement, the guided filtering [24] is applied to obtain the
Our two-layer model is not simply substituting the feature final transmission [Fig. 6(e)]. The guided filter is simple and
vector at the current superpixel in (24) with an augmented effective, producing good results in less time compared with
one by the features of eight surrounding superpixels, though the soft matting in [2].
the transmission estimation on our model is merged into
one Gaussian in (29). The underlying assumption of using D. Subsequent Processing for Haze Removal
the augmented feature vector in (24) is the independence
between any transmissions, given the features of the current We complete the major work for haze removal once we
superpixel and its neighbors. This independent assumption obtain the transmission, except for the unknown atmosphere
would neglect the interactions between neighboring transmis- light A. In [7], A is considered as the value of input image
sions, e.g., structural consistency of neighboring transmissions. corresponding to the top 0.1% brightest pixels of dark channel
In contrast, the second GPR layer in our model encodes these with corrections due to the multiple scattering effect [29].
interactions. In this paper, we select the value of an input image corre-
Equation (30) shows that the final estimation m(t ∗ ) of sponding to the top 0.1% darkest pixels of transmission as A,
the transmission at a superpixel is a linear combination because our estimation is more precise than the simple dark
of the predicted estimates from the two layers, weighted channel. Meanwhile, the correction method is kept
by their prediction variances. This two-layer model takes A∗ = (1 − η)A (32)
the advantage of the GP regression whereby the prediction
uncertainty is provided in terms of the prediction variance. where η is a correction parameter, which is commonly chosen
The final transmission relies more on the prediction of a GPR as 0.02, 0.01, or 0.
layer with the lower variance giving a more accurate estimate. For the sky region, we use the compensation in [7]
Consequently, the transmission estimation of the two-layer ti ( pi ) = 2tα − ti ( pi ) (33)
combination, as shown in Fig. 6(d), attenuates ringing effects
appearing in the output of the first GPR layer [Fig. 6(b)], where tα is the transmission boundary of the sky and the
and meanwhile preserves more details than the transmission scenery.
map generated by the second layer [Fig. 6(c)]. It is also Finally, we manually set Dvis j and Dvisi with
worth noting from (31) that the variance σ 2 (t ∗ ) is smaller Dvis j  Dvisi to remove haze, and as shown in Fig. 6(f), the
than that of the prediction from either layer. This decreased restored image is almost the same as the original [Fig. 6(g)].
2512 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

TABLE I
M EAN VARIANCES (×10−6 ) FOR F OUR H AZY I MAGES

every image is always the lowest among the three. Both Fig. 7
and Table I show that the two-layer GPR model is able to
generate more accurate transmission estimation, upon which
high-quality dehazing images can be restored. In the following
experiments, we present the dehazing performance of the two-
layer GPR for the comparisons with the existing dehazing
Fig. 7. Variance distributions of the first (feature to transmission), second
(neighbor to transmission), and two-layer GPR. methods.

B. Comparisons With Other Haze Removal Methods


V. E XPERIMENTAL R ESULTS AND A NALYSIS
Our comparisons begin with the experiments on synthetic
In this section, we first validate the use of the two-layer
hazy images in which the ground truth haze-free images are
GPR model over one layer ones by comparing their estimation
available. Original images and their corresponding depth maps
variances. Second, we show our experimental results and com-
are used to simulate haze images based on the work by
parisons with the dehazing existing methods [2], [10], [11],
Saxena et al. [30]. We compare our haze removal results of
and perform both qualitative and quantitative comparisons on
these synthetic images with the existing methods. Fig. 8 shows
synthetic and real-world hazy images. Finally, we use different
the haze-free images, the simulated hazy images, and the haze
groups of training examples to train the regressor in order
removal results of different methods on the synthetic images.
to show the effectiveness of the proposed screening process.
In the comparisons, the work by He et al. [2] overestimates the
In the experiment, we use ten images as the training set.1
transmission; therefore, the whole image is dim. The results
A. Transmission Variance Comparisons also have color distortions and fail to deal with sunny and
white scenes. Tang et al.’s results [10] also produce color
The transmission estimates from the first (feature to trans- distortions and generate dark images due to overdehazing
mission), second (neighbor to transmission), and two-layer
caused by inaccurate transmission estimation. Zhu et al. [11]
models are given by (25), (27), and (28), respectively.
underestimate the haze. There is a large amount of haze that
As shown in Fig. 6(d), the transmission map estimated from remained in the image, as shown in the second row. By
the two-layer model combines the details predicted from the
contrast, our results have a few color distortions. The haze
first layer [Fig. 6(b)] and the smoothness from the second removal results restore to the previous haze-free images ,
one [Fig. 6(c)], showing the merits of the weighted mean esti- which are the best among the four methods. The original colors
mation (30) for the two-layer model. Herein, we compare the
of the buildings and the roads are well preserved. Compared
variances of these estimates to experimentally validate (31). with the other methods, our results look the most natural.
We take 660 superpixels of four hazy images for analysis, and
We also collect real-world hazy images and remove the haze
obtain three groups of the 660 transmission variances at these
from them with the existing methods in [2], [10], and [11].
superpixels from the first, second, and two-layer models. The Fig. 9 shows our results and the results of the other methods.
normalized (by 660) distributions or histograms of the three
He et al.’s method generates dark images due to the overde-
variances are shown in Fig. 7. The y-axis shows the fraction of hazing of DCP, and it fails on sky regions, as shown in the
the variance values that fall within the bins of the x-axis. We third row of Fig. 9. Zhu et al.’s results have haze in the images.
can see that most of transmission variances of the two-layer
Tang et al.’s results perform better than the two methods but
model are smaller than those variances of the first and second also have some degree of color distortions. He et al.’s and
ones. Over 25% of variances from the two-layer model are
Tanget al.’s methods tend to overestimate the thickness of
smaller than 0.01, while about 10% from the first and second
haze (or underestimate the transmission), which results in dim
ones fall within this range. We also take the mean of the three scenes and different degrees of color distortions. By contrast,
groups of variances for four hazy images, as shown in Table I.
our results have little color distortions and look more realistic.
The mean value of the variances for 660 superpixels of the
The truck is clearer than the other results as shown in the
second layer is higher than the first one, as the prediction second row of Fig. 9. The global consistency and the details
from the second layer is given upon the output of the first
are well preserved.
one, having more uncertainties. In addition, the prediction To evaluate the results in an objective way, we use two
from features might be more informative than the smoothness.
metrics with reference for image quality assessment, i.e.,
As expected, the mean value of the two-layer GPR model for
the peak signal-to-noise ratio (PSNR) [31] and the structural
1 The source code and testing images can be found at similarity (SSIM) [32]. We simulate 30 hazy images, and
https://github.com/dlut-dimt/Two-Layer-GPR-Dehazing. employ the box plots to visualize the statistics of dehazing
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2513

Fig. 8. Visual comparisons on synthetic hazy images. (a) Haze-free images. (b) Hazy images. (c) He et al.’s results [2]. (d) Tang et al.’s results [10].
(e) Zhu et al.’s results [11]. (f) Our results.

Fig. 9. Visual comparisons on real-world hazy images. From left to right: hazy images, He et al.’s results [2], Tang et al.’s results [10],
Zhu et al.’s results [11], and our results.

results evaluated by the two metrics. Fig. 10 shows the box of the boxes show the extent of the data falling out of the
plots of SSIM and PSNR of different methods on the synthetic quartile. The higher median values of PSNR and SSIM, the
images. The blue horizontal lines indicate the lower and upper results are closer to their corresponding haze-free images.
quartile values. The red horizontal lines in the box denote In the comparison of SSIM and PSNR, the median values
the median values. The whiskers extending from two ends of He et al.’s results are lower than the other methods, as
2514 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

Fig. 10. Box plot of quantitative comparisons on synthetic hazy images. (a) SSIM of different methods. (b) PSNR of different methods.

Fig. 11. Box plot of quantitative comparisons on real-world hazy images. (a) Blur metric of different methods. (b) BIQI of different methods.

a large portion of white regions appear in these images where The BIQI metric computes blockness, blur, corner outliers,
DCP does not work very well. By contrast, our results have and noise of an image, and combines them into a quality
the highest median values of SSIM and PSNR among the four score in order to evaluate the feature distortions. The quality
methods. score has a value between 0 and 100. Lower values stand for
For quantitative evaluation on haze removal results, we smaller distortions or higher quality. As shown in Fig. 11(b),
use 18 real-world hazy images for experiments. It is still an the median values of these methods have no significant gaps.
open issue to objectively evaluate the dehazing quality without Among them, our results output the lowest median value
any reference [33]. We use two popular metrics, i.e., the blur of BIQI, since the second GPR layer is able to smear structural
metric [34] and blind image quality indices (BIQI) [35] for inconsistency. Fig. 11 shows that our method produces the
our comparison. Fig. 11 shows the box plots of the image dehazed images of high quality under these two metrics.
qualities on real-world hazy images. The blur metric evaluates However, it is still under investigation for the community to
the image quality in terms of blur perception, and the blur devise a comprehensive metric, which reflects both sharpness
metric ranges from 0 to 1, indicating the best to the worst. and structural information, for the evaluation of dehazing
As shown in Fig. 11(a), the median value of Tang et al.’s and algorithms. Though it is quite interesting, we have to leave this
Zhu et al.’s results is almost the same and is higher than the comprehensive objective comparisons for future work that is
another two methods, which means that their results have the beyond the scope of this paper. To be more intuitive, we show
worst image quality from the blur aspect. These results can be the mean values of SSIM, PSNR, blur, and BIQI of different
partially explained by the simulated training images used in methods in Table II.
their regression. This discrepancy between training and testing To make a subjective verification on visual effect, we also
examples generates inaccurate transmission prediction for real- make a perceptual comparison [33] with the existing methods
world natural images. Our results have comparable median in [2], [10], and [11]. We establish a test database with
value with those of the He et al.’s but show a significantly 18 real-world hazy images and dehazed results of different
smaller lower quartile value. We are able to produce quite dehazing methods. We asked 22 young observers (12 males
sharp images in some cases. and 10 females) to rate the resultant images of different
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2515

TABLE IV
M EAN PSNR OF D IFFERENT G ROUPS OF T RAINING E XAMPLES

the highest mean value and the lowest std. The high and
stable performance can be explained, as the transmissions of
natural scenes are well characterized by our two-layer model.
Meanwhile, our screening process is able to find suitable
examples to train the model. Both the model and the screening
exhibit the consistency with observers’ perceptual assessments.
Fig. 12. Bar plots of the MOS values of different dehazing methods.
C. Training Examples With Various Haze Concentrations
TABLE II
In Section III, we develop a two-step screening process
M EAN VALUES OF SSIM, PSNR, BLUR, AND BIQI OF
D IFFERENT M ETHODS
to choose training images. We use this process to group the
images in the training image data set into three categories
according to the haze concentration: 1) images without haze;
2) images with dense haze; and 3) images with moderate
haze. We then use the images in these three groups to train
our two-layer GPR models, and generate dehazing results in
Fig. 13 with the trained model from different groups of training
TABLE III examples. The GP regressors trained from images without haze
M EAN AND S TANDARD D EVIATION OF MOSs OF or dense haze underestimate the transmission, and there still
D IFFERENT D EHAZING M ETHODS
exists notable haze in the resultant images. In addition, the
sky regions also have color distortions. We can see that the
results from the training images with moderate haze perform
the best among the three groups. The restored results retain the
details of the images and have high brightness and contrast.
dehazing methods with a value from 1 to 10. Higher values The scenes are also restored well. This shows the importance
indicate better perceptual quality. The dehazed results were of selecting suitable images for training and that our proposed
presented to the observers in a random order. Hazy images also screening process works well. We also use the nonreference
need to be scored as references. All these observers have no blur metric [34] to perform the quantitative evaluation on
background nor prior knowledge on dehazing algorithms. We haze removal results of different training groups, as shown
compute the mean opinion score (MOS) for all testing images in Fig. 14 where box plots are given.2 The three groups have
and compare the performance of dehazing methods in terms different degrees of blurring on the results, indicating that the
of the mean and standard deviation (std) of the MOS values haze removal results are affected by training examples. Among
for the outputs of these methods. The mean value provides the three groups, training images without haze generate a
the overall MOS of a method, while the std value indicates model that is poor for sharpening hazy images. The lower
the stability of the method. Fig. 12 shows the bar plots of the median values on the blur metric from the second and third
MOS of different dehazing methods, where the height of a bar groups show that the inclusion of hazy images is definitely able
indicates that the mean and the height of the vertical line on to improve the ability of the learned model for prediction. This
top of the bar denote the standard deviation. In addition, the validates the effectiveness of the first screening stage. More-
values of mean and standard deviation of the MOSs are shown over, the performance can be further improved by filtering
in Table III. out images with dense haze with the second screening. The
As shown in Fig. 12, observers have consensus on the qual- lowest median value and the upper quartile value verify the
ity improvements of all dehazing methods over hazy images. effectiveness and stability of our screening process.
Zhu et al.’s method relies on the atmospheric scattering model In our experiment, we also simulate hazy images with
that tends to underestimate the transmission for dense hazy different degrees of haze on haze-free images and use the
images, leading to the lowest mean MOS among the compared three groups of training examples above to remove haze from
methods. Tang et al.’s and He et al.’s methods share similar synthetic hazy images. We group the simulated hazy image
mean scores, but He et al.’s method has a higher std value. into three categories: 1) images with thin haze; 2) images
He et al.’s method based on DCP does not work well on some with dense haze; and 3) images with moderate haze. Table IV
images with large white or sky regions. Tang et al.’s method is shows the mean value of PSNR of different groups of training
able to output more stable performance because of the learned 2 All the testing and resultant images for the comparisons in
model from a variety of natural scenes. Our method produces Figs. 10, 11, and 14 are available in the Supplemental Materials.
2516 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 27, NO. 12, DECEMBER 2017

Fig. 13. Comparisons among the different groups of training examples. From left to right: hazy images, haze removal results of the first group, haze removal
results of the second group, and haze removal results of the third group.

is expected to improve the efficiency of the training; 2) the


addition of other features may improve the precision of the
regression further; and 3) the regression framework can be
used to handle other related problems, like the underwater
scene [37] or night image [38] enhancement.
ACKNOWLEDGMENT
The authors would like to thank Prof. M.-T. Sun at the
University of Washington for his valuable suggestions and
carefully proofreading this paper. They would also like to
thank Dr. J. Wang and Dr. J. Yang for providing the results of
their method [10] on all the synthetic and real-world images
in this paper.
Fig. 14. Box plot of blur metric comparisons among different groups of
training examples. R EFERENCES
[1] R. Fattal, “Single image dehazing,” ACM Trans. Graph., vol. 27, no. 3,
examples for hazy images with different haze concentrations. p. 72, Aug. 2008.
[2] K. He, J. Sun, and X. Tang, “Single image haze removal using dark
The mean value of PSNR evaluation of the third training channel prior,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12,
group is always higher than the other two training groups, and pp. 2341–2353, Dec. 2011.
the applicability of the selected training examples is proved. [3] R. T. Tan, “Visibility in bad weather from a single image,” in Proc.
IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2008, pp. 1–8.
Therefore, screening appropriate training images is important [4] J.-P. Tarel and N. Hautière, “Fast visibility restoration from a single
and can greatly increase the accuracy and efficiency of the color or gray level image,” in Proc. IEEE 12th Int. Conf. Comput. Vis.,
dehazing method. Sep./Oct. 2009, pp. 2201–2208.
[5] K. B. Gibson, D. T. Vo, and T. Q. Nguyen, “An investigation of dehazing
effects on image and video coding,” IEEE Trans. Image Process., vol. 21,
VI. C ONCLUSION AND D ISCUSSION no. 2, pp. 662–673, Feb. 2012.
[6] Y.-H. Lai, Y.-L. Chen, C.-J. Chiou, and C.-T. Hsu, “Single-image
In this paper, we demonstrate a two-layer GPR model to dehazing via optimal transmission map under scene priors,” IEEE Trans.
estimate the transmission and remove haze. By using training Circuit Syst. Video Technol., vol. 25, no. 1, pp. 1–14, Jan. 2015.
[7] X. Fan, Y. Wang, R. Gao, and Z. Luo, “Haze editing with natural
examples, the two-layer GPR establishes direct relationships transmission,” Vis. Comput., vol. 23, no. 1, pp. 137–147, 2016.
from the input image to the depth-dependent transmission, [8] H. Huang and N. Wu, “Fast facial image super-resolution via local linear
and meanwhile learns local image priors to further improve transformations for resource-limited applications,” IEEE Trans. Circuits
Syst. Video Technol., vol. 21, no. 10, pp. 1363–1377, Oct. 2011.
the estimation. We also provide a method to collect training [9] H. He and W.-C. Siu, “Single image super-resolution using Gaussian
pairs for images of natural scenes. Experimental results show process regression,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.,
the advantage of our algorithm compared with other existing Jun. 2011, pp. 449–456.
[10] K. Tang, J. Yang, and J. Wang, “Investigating haze-relevant features in a
methods. Our model still has room for improvement and learning framework for image dehazing,” in Proc. IEEE Conf. Comput.
extension: 1) a faster algorithm for online learning of GPs [36] Vis. Pattern Recognit., Jun. 2014, pp. 2995–3002.
FAN et al.: TWO-LAYER GPR WITH EXAMPLE SELECTION FOR IMAGE DEHAZING 2517

[11] Q. Zhu, J. Mai, and L. Shao, “A fast single image haze removal algorithm [38] X. Jiang, H. Yao, S. Zhang, X. Lu, and W. Zeng, “Night video
using color attenuation prior,” IEEE Trans. Image Process., vol. 24, enhancement using improved dark channel prior,” in Proc. IEEE Int.
no. 11, pp. 3522–3533, Nov. 2015. Conf. Image Process., Sep. 2013, pp. 553–557.
[12] K. B. Gibson, S. J. Belongie, and T. Q. Nguyen, “Example based
depth from fog,” in Proc. IEEE Int. Conf. Image Process., Sep. 2013,
pp. 728–732.
[13] X. Ren and J. Malik, “Learning a classification model for segmentation,”
in Proc. IEEE Int. Conf. Comput. Vis., Oct. 2003, pp. 10–17. Xin Fan (M’06) was born in 1977. He received the
[14] L. Chen, G. Lu, and D. Zhang, “Effects of different Gabor filters B.E. and Ph.D. degrees in information and commu-
parameters on image retrieval by texture,” in Proc. 10th Int. Multimedia nication engineering from Xi’an Jiaotong University,
Modelling Conf. (MMM), 2004, pp. 273–278. Xi’an, China, in 1998 and 2004, respectively.
[15] W. T. Freeman, E. C. Pasztor, and O. T. Carmichael, “Learning low-level From 2006 to 2007, he was with Oklahoma State
vision,” Int. J. Comput. Vis., vol. 40, no. 1, pp. 25–47, 2000. University, Stillwater, OK, USA, as a Post-Doctoral
[16] X. Zhao, S. Wang, S. Li, and J. Li, “Passive image-splicing detection Research Fellow. In 2008, he was with the Univer-
by a 2-D noncausal Markov model,” IEEE Trans. Circuits Syst. Video sity of Texas Southwestern Medical Center, Dallas,
Technol., vol. 25, no. 2, pp. 185–199, Feb. 2015. TX, USA, for the second post-doctoral training. He
[17] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine joined the School of Software, Dalian University of
Learning. Cambridge, MA, USA: MIT Press, 2006. Technology (DUT), Dalian, China, in 2009, where
[18] C. M. Bishop, Pattern Recognition and Machine Learning. New York, he is currently a Full Professor and Vice Dean of the School for Research.
NY, USA: Springer, 2006. He is also the Duty Dean of the DUT-RU International School of Information
[19] P. Domingos, “A few useful things to know about machine learning,” and Science Technology, Dalian. His research interests include computational
Commun. ACM, vol. 55, no. 10, pp. 78–87, 2012. geometry and machine learning, and their applications to low-level image
[20] C. R. González and Y. S. Abu-Mostafa, “Mismatched training and test processing and DTI-MR image analysis.
distributions can outperform matched ones,” Neural Comput., vol. 27,
no. 2, pp. 365–387, 2014.
[21] N. Cristianini and J. Shawe-Taylor, Support Vector Machines and Other
Kernel-Based Learning Methods. Cambridge, U.K.: Cambridge Univ. Yi Wang received the B.E. and Ph.D. degrees in
Press, 2000. computer science from Jilin University, Changchun,
[22] X. Fan, R. Gao, and Y. Wang, “Example-based haze removal with two- China, in 2002 and 2009, respectively.
layer Gaussian process regressions,” in Pacific Graphics Short Papers. She has been a Lecturer with the School
Geneve, Switzerland: The Eurographics Association, 2014. of Software, Dalian University of Technology,
[23] C. O. Ancuti, C. Ancuti, C. Hermans, and P. Bekaert, “A fast semi- Dalian, China, since 2009. Her research interests
inverse approach to detect and remove the haze from a single image,” include machine learning, computer vision, and
in Proc. 10th Asian Conf. Comput. Vis., 2011, pp. 501–514. virtual reality.
[24] K. He, J. Sun, and X. Tang, “Guided image filtering,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 35, no. 6, pp. 1397–1409, Jun. 2013.
[25] F. Cao, M. Cai, and Y. Tan, “Image interpolation via low-rank matrix
completion and recovery,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 25, no. 8, pp. 1261–1270, Aug. 2015.
[26] J. S. Yedidia, W. T. Freeman, and Y. Weiss, “Understanding belief Xianxuan Tang was born in 1994. She received
propagation and its generalizations,” in Exploring Artificial Intelligence the B.E. degree in software engineering from Dalian
in the New Millennium, vol. 8. San Francisco, CA, USA: Morgan Maritime University, Dalian, China, in 2014. She is
Kaufmann, 2003, pp. 236–239. currently pursuing the master’s degree with Dalian
[27] E. B. Sudderth, A. T. Ihler, W. T. Freeman, and A. S. Willsky, University of Technology, Dalian.
“Nonparametric belief propagation,” in Proc. IEEE Conf. Comput. Vis. Her research interests include computer vision,
Pattern Recognit., Jun. 2003, pp. 605–612. image processing, and machine learning.
[28] M. D. Springer and W. E. Thompson, “The distribution of products of
beta, gamma and Gaussian random variables,” SIAM J. Appl. Math., vol.
18, no. 4, pp. 721–737, 1970.
[29] T. Nishita, Y. Dobashi, K. Kaneda, and H. Yamashita, “Display method
of the sky color taking into account multiple scattering,” in Proc. Pacific
Conf. Comput. Graph. Appl., 1996, pp. 117–132.
[30] A. Saxena, M. Sun, and A. Y. Ng, “Learning 3-D scene structure from a
single still image,” in Proc. IEEE 11th Int. Conf. Comput. Vis., Oct. 2007, Renjie Gao received the B.E. and M.S. degrees
pp. 1–8. in software engineering from Dalian University
[31] Q. Huynh-Thu and M. Ghanbari, “Scope of validity of PSNR in of Technology, Dalian, China, in 2012 and 2015,
image/video quality assessment,” Electron. Lett., vol. 44, no. 13, respectively.
pp. 800–801, Jun. 2008. He is currently with China Construction Bank,
[32] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image Beijing, China. His research interests include com-
quality assessment: From error visibility to structural similarity,” IEEE puter vision, image processing, and machine learn-
Trans. Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004. ing.
[33] K. Ma, W. Liu, and Z. Wang, “Perceptual evaluation of single
image dehazing algorithms,” in Proc. IEEE Int. Conf. Image Process.,
Sep. 2015, pp. 3600–3604.
[34] F. Crete, T. Dolmiere, P. Ladret, and M. Nicolas, “The blur effect:
Perception and estimation with a new no-reference perceptual blur
metric,” Proc. SPIE, vol. 6492, p. 64920I, Feb. 2007. Zhongxuan Luo received the B.S. and M.S. degrees
[35] A. K. Moorthy and A. C. Bovik, “A two-step framework for constructing from Jilin University, Changchun, China, in 1985
blind image quality indices,” IEEE Signal Process. Lett., vol. 17, no. 5, and 1988, respectively, and the Ph.D. degree from
pp. 513–516, May 2010. Dalian University of Technology, Dalian, China, in
[36] Y. Kwon, K. I. Kim, J. Tompkin, J. H. Kim, and C. Theobalt, “Efficient 1991, all in computational mathematics.
learning of image super-resolution and compression artifact removal with Since 1997, he has been a Full Professor with the
semi-local Gaussian processes,” IEEE Trans. Pattern Anal. Mach. Intell., School of Mathematical Sciences, Dalian University
vol. 37, no. 9, pp. 1792–1805, Sep. 2015. of Technology, where he is currently the Dean of the
[37] H. Lu, Y. Li, and S. Serikawa, “Underwater image enhancement using School of Software. His research interests include
guided trigonometric bilateral filter and fast automatic color correction,” computational geometry and computer vision.
in Proc. IEEE Int. Conf. Image Process., Sep. 2013, pp. 3412–3416.

You might also like