Comparing Color Descriptors between Image Segments for Saliency

Detection

Anurag Singh

, Henry Chu

1,2

and Michael Pratt

The Center for Advanced Computer Studies, University of Louisiana at Lafayette, Lafayette, LA, U.S.A.

Informatics Research Institute, University of Louisiana at Lafayette, Lafayette, LA, U.S.A.

W. H. Hall Department of Electrical & Computer Engineering, University of Louisiana at Lafayette, Lafayette, LA, U.S.A.

Keywords:

Visual Saliency, Superpixels, Earth Mover’s Distance, Dominant Color Descriptors.

Abstract:

Detecting salient regions in an image or video frame is an important step in early vision and image understand-

ing. We present a visual saliency detection method by measuring the difference in color content of an image

segment with that of its neighbors. We represent each segment with richer color descriptors in the form of a

regional dominant color descriptor. The color difference between a pair of neighbors is found using the Earth

Mover’s Distance. The cost of moving color descriptors between neighboring segments robustly captures the

difference between neighboring segments. We evaluate our method on standard datasets and compare it with

other state-of-the-art methods to demonstrate that it has better true positive rate at a ﬁxed false positive rate

in detecting salient pixels relative to the ground truth. The proposed method uses local cues without being an

edge highlighter, a common problem of local contrast-based methods.

1 INTRODUCTION

Visual saliency detection is a useful early step in many

computer vision applications to focus attention to the

most prominent object or region in an image or video.

As such, a salient region is one that stands out, or acts

as an outlier, with respect to other regions in its vicin-

ity. The important considerations for a pixels or re-

gion to be salient are as follows. First, we consider

the uniqueness in color space: color is one of the most

important feature which differentiates between salient

and non salient pixels; a salient group often stands

out because it is different in color content compared

to other groups. Secondly, visual ﬁxation occurs in

clusters (Treisman, 1982), (Ben-Av et al., 1992). The

so-called superpixels are a suitable representation of

clustering of pixels. Superpixels are mostly homo-

geneous in image content and often conform to edge

deﬁnition in an image. Hence, superpixel segment

(Felzenszwalb and Huttenlocher, 2004) can be chosen

as the lowest level primitives. When so doing, it has

the additional beneﬁt of reducing the computational

cost of comparisons drastically when compared to the

cost using pixels or pixel-centered neighborhoods.

A number of methods have been proposed in

the literature to detect image saliency (Liu et al.,

2007),(Itti et al., 1998),(Harel et al., 2006),(Ma and

Zhang, 2003),(Goferman et al., 2010),(Lin et al.,

2013). The color contrast-based methods for saliency

detection is done using either global comparison

(Singh et al., 2014), (Cheng et al., 2011), (Perazzi

et al., 2012) or local comparison (Goferman et al.,

2010), (Liu et al., 2007). Since the eye ﬁxation pro-

cess of biological vision responses to local cues, the

local comparison methods are de facto edge high-

lighters in that they highlight the region around the

edge boundary (Fig 1).

In this paper we present a method which uses lo-

cal contrast to highlight a whole salient segment. The

saliency detection method is based on the difference

of the salient region within its neighborhood, or the

outlierness of the region. We summarize the color

content of each segment by its dominant color de-

scriptors (DCD). An advantage of using the DCDs is

that DCDs of all image segments or regions are of the

same size, irrespective of the number of pixels in the

region. This DCD property facilitates the use of the

Earth Mover’s Distance to measure the difference of

the region’s DCD and those of its neighbors to deter-

mine the outlierness of a region.

The novelty of our method lies in exploiting the

segment signatures of DCDs and the choice of met-

ric, viz. the Earth Mover’s Distance (EMD) to com-

pute the color contrast. EMD was considered to be the

558

Singh, A., Chu, H. and Pratt, M.

Comparing Color Descriptors between Image Segments for Saliency Detection.

DOI: 10.5220/0005667705580565

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 558-565

ISBN: 978-989-758-173-1

Figure 1: Examples of image saliency detection. Top row:

Input image. Middle row, left to right: Methods SO (Liu

et al., 2007), IT (Itti et al., 1998), GB (Harel et al., 2006);

Bottom row, left to right: Methods MZ (Ma and Zhang,

2003), CA (Goferman et al., 2010), EMD proposed in the

present paper. The saliency map’s highlighted area varies

depending on the method used.

best metric when comparing equal sized signatures,

such as in the case of DCDs (Rubner et al., 2000a).

EMD also enables us to make a weighted comparison

between color feature vectors, so that color is used

proportional to its importance. The contrast differ-

ence between segments is estimated by the amount of

work needed to move the color from one segment to

another. The more the amount of work is required for

moving the contents, the more is the perpetual differ-

ence between segments.

When compared to the other color-contrast meth-

ods, the proposed algorithm uses only the neighbor-

ing segments to infer saliency and uses richer descrip-

tors than, say, color averages. By highlighting whole

salient objects, it overcomes problems associated with

the local methods that act as edge highlighters (Per-

azzi et al., 2012).

2 SALIENCY DETECTION

We formulate the visual saliency as a measure of the

amount of color difference between neighboring im-

age segments. This formulation results in a graphi-

cal structure where each image segment is seen as the

receiver of color from connected segments. In this

section we discuss in detail the color description of

each segment, a metric for comparison and a multi-

resolution step for ﬁnding best segment. Fig. 2 shows

the ﬂow diagram for the saliency detection algorithm.

2.1 Superpixel Segmentation and

Summarization

An input image is ﬁrst divided into superpixel seg-

ments. Superpixels are usually homogeneous and

conform to image edges, making them good primi-

tives for comparison. Superpixels are of various sizes

and to make true comparisons we need to use a nor-

malized representation. Each superpixel’s color infor-

mation is summarized using the DCD.

The dominant color descriptor was introduced in

the MPEG-7 standard (Manjunath et al., 2001) to pro-

vide a compact description of an image for appli-

cations such as content-based image retrieval from

a collection of images of different sizes. We com-

pute a DCD not for the entire image but for each

superpixel segment. A descriptor for each super-

pixel therefore consists of a set of representative col-

ors and their corresponding percentages in the super-

pixel region. DCDs provide a perceptually closer rep-

resentation of the input image (Singh et al., 2014),

hence they form a better descriptor model for com-

parison. The DCD for the kth image segment is given

by DCD

= {(c

, p

) : i = 1, ··· ,N

}, where c

is a

color in the CIE L*a*b* space, p

is a percentage

of pixels in the superpixel represented by the corre-

sponding color. The CIE L*a*b* color space is cho-

sen, as it supports double opponency and is perceptu-

ally similar to the color scheme in the human visual

cortex (Engel et al., 1997). There are different ways

proposed in the literature to compare two DCDs. One

such comparison method (Yang et al., 2008) is given

(DCD

,DC D

) = 1−

∑

i=1

∑

j=1

{(1 − (p

− p

2 j

))

× min(p

, p

2 j

1i,2 j

}

(1)

where a

1i,2 j

is the similarity between colors c

and

2 j

, given by

1i,2 j



1 − d

1i,2 j

, d

1i,2 j

≤ T

0, otherwise

(2)

where d

1i,2 j

is the Euclidean distance between colors

and c

2 j

, and T

is a threshold set between 0 and 1.

2.2 Saliency Computation

In this section we discuss in detail the steps followed

to compute the saliency of a superpixel segment. As

mentioned before, the color signature of each segment

is represented by the clustered dominant colors.

Comparing Color Descriptors between Image Segments for Saliency Detection

559

Input Segments DCD Segment Signatures EMD Flow Final Output

Figure 2: Flow diagram for the proposed local comparison-based saliency detection algorithm.

Algorithm 1: DCD computation.

Input: Pixels ∈ Superpixel

Output: Dominant Color Desciptors

RGB → CIEL ∗ a ∗ b∗ transform;

N = superpixel.size() ;

while i ≤ N do

Read all pixels in a superpixel[i] ;

Create 4 DCD cluster C

;

Cluster colors using K-means;

for j ← 1; j < 4 do

for k ← j + 1; j < 4 do

if kC

−C

< δ then

Merge Clusters ;

else

Continue;

Compute % of pixels for each DCD;

Figure 3: Image signature computed using dominant color

descriptors.

The difference in DCD signatures is calculated us-

ing the Earth Mover’s Distance (EMD). The EMD

measures the minimum amount of work required to

move the color contents of one superpixel segment

to those of another superpixel segment. This metric

gives us the amount of perceptual color difference be-

tween the superpixel segments.

The EMD ﬁnds its roots in a known solution of the

transportation problem or Monge-Kantorovich prob-

lem. Suppose there are several suppliers and several

consumers; the solution of the transportation problem

is to ﬁnd the least expensive ﬂow of goods from the

suppliers to the consumers.

For comparing visual information, the EMD

(Rubner et al., 2000a) was introduced as a metric to

compare two distributions in a transformed feature

space, where a measure is computed as the amount of

ground distance moved. EMD has been used as a met-

ric for comparing the dissimilarity between two distri-

butions of color and texture in image retrieval (Rubner

et al., 2000b) and saliency detection (Lin et al., 2013).

The EMD is formalized as a linear program-

ming problem (Rubner et al., 2000a). In our con-

text, let P and Q be two segments represented by

the dominant color descriptors as two signatures

DCD

= ((c

)· ·· (c

)) and DCD

((c

)· ·· (c

)), where c

and c

Q j

are

color cluster representatives of P and Q, respectively,

and w

Q j

are the corresponding percentages of

pixels as weight of the clusters. Let D = [d

i j

] be the

cost matrix or the ground distance matrix for moving

color between DCD

and DCD

so that d

i j

is the pair-

wise distance in the CIE L*a*b* space between two

colors, one from each DCD:

i j

= ||c

− c

Q j

||. (3)

The ﬂow F = [ f

i j

] between DCD

and DCD

is min-

imized by

WORK(P,Q, F) =

∑

i=1

∑

j=1

i j

. (4)

The EMD is deﬁned as the work done from Equation

4, normalized by the total ﬂow and given as

EMD(P, Q) =

∑

i=1

∑

j=1

i j

∑

i=1

∑

j=1

i j

(5)

where total ﬂow

∑

i=1

∑

j=1

i j

is given as

min(

∑

i=1

∑

j=1

Q j

The EMD solution is subjected to the following

constraints:

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

560

1. Supplies are moved only from P to Q one way;

i.e.,

i j

≥ 0,1 ≤ i ≤ m,1 ≤ j ≤ n

2. Limits the amount of supplies that can be sent

from clusters to the weights; i.e.,

∑

j=1

i j

≤ w

,1 ≤ i ≤ m

3. Limits the clusters in their weights; i.e.,

∑

i=1

i j

≤ w

,1 ≤ j ≤ n

4. Force to move maximum amount of supply; i.e.,

∑

i=1

∑

j=1

i j

= min(

∑

i=1

∑

j=1

Q j

)

The major advantage of using EMD for superpixel

comparison is that it captures the signature of super-

pixels better than an average color or position value.

Another advantage of using EMD is that it matches

similarities (or ﬁnds dissimilarities) in a way that is

consistent with human perception. EMD is a true met-

ric of comparing the distributions of similar masses,

which is the case when the superpixels are approxi-

mately similarly sized. Our use of the EMD is similar

to comparing histograms, where the amount of work

done to move the bin contents from one histogram to

another is a measure of the dissimilarity between the

two histograms.

We note that EMD has been used to compare his-

tograms in saliency detection (Lin et al., 2013). In that

work, EMD was used to compute the center-surround

difference in the framework of (Itti et al., 1998). The

center-surround difference is computed directly from

the color values of two image regions, so that the his-

tograms have a large number of bins. Two histograms

a and b have the same number of bins and the corre-

sponding bins a

and b

represent the same range of

colors, for all i. In our work the color information in

each segment is summarized by a DCD. Each of the

DCDs being compared may have up to 4 components

and each component can represent a different color. In

terms of EMD computation, the cost matrix D in (Lin

et al., 2013) is large, computed once, and ﬁxed for all

pairs of histograms in an image whereas the matrix D

in our work is much smaller (at most 4 by 4) and is

computed for each pair of DCDs being compared.

Aggregating the Difference: The saliency of the

Pth segment is computed by aggregating all the work

done associated with moving colors from neighbor-

ing segments. The aggregated average saliency for

segment P is given by

Sal

∑

k=1

EMD(P, Q

) (6)

where K is the total number of neighbors, and the

EMD between segment P and a neighboring segment

is found using Equation 5.

Biasing for Center: There is an inherent central

bias in visual saliency maps (Borji and Itti, 2013).

The central bias is updated by penalizing the segments

that are far from the center in the following way:

Sal

(1 − δ) (7)

where

Sal

is the saliency value for Segment P and δ

is the normalized distance of the segment center from

the center of the image.

2.3 Normalization Step

The human visual system suppresses the low level re-

sponses and focuses attention on stimuli with higher

level responses. In this algorithm, the focus is desired

to be on salient objects. A generic objectness measure

is used, which is the probability of occurrence of an

object in a window (Alexe et al., 2012). Sampling for

object windows gives the notion of objectness (Sun

and Ling, 2013), which ensures a higher probability

value for the occurrence of an object. The normalized

saliency map is given by

Sal

norm

(Sal

map

+ Ob j

map

) (8)

where S al

map

is the ﬁnal saliency map formed by

compositing the saliency values of the superpixel seg-

ments computed from Equation 7 and Ob j

map

is the

objectness map.

2.3.1 Choosing the Best Segmented Image

The major limiting factor is the quality of the super-

pixel segmentation algorithm. If the segmentation al-

gorithm divides the salient region into very small seg-

ments, the overall saliency algorithm enhances small

segments found between salient and non-salient re-

gions. In Fig. 4, an example of this is shown. To

Input Image Segments Saliency Map Ground Truth

Figure 4: Limitations due to non cohesive segmentation.

Non-cohesive(top row) and cohesive (bottom row) segmen-

tation.

Comparing Color Descriptors between Image Segments for Saliency Detection

561

Input Saliency Map Ground Truth Input Saliency Map Ground Truth Input Saliency Map Ground Truth

Figure 5: Saliency map results for four publicly available, standard datasets. Each row corresponds to a different dataset. On

each row, we show three representative triplets of an input image, the saliency map, and the ground truth image.

overcome this, a measure for ﬁnding best segment

(Bagon et al., 2008) is integrated into the algorithm.

The goodness measure is based on the following con-

ditions: i) every pixel belongs to a region, ii) every

region is spatially connected, iii) regions are disjoint;

and iv) all pixels in a region satisfy a speciﬁed simi-

larity.

Initially, the input image is segmented into multi-

ple images, each with superpixels at a particular res-

olution. The resolution that gives the the best good-

ness measure is chosen for further processing. From

this resolution, the superpixels image with best cohe-

sive segmentation is generated. The more cohesive

the segments will result in better saliency detection.

3 EXPERIMENTS

3.1 Implementations

The local comparisons based saliency detection algo-

rithm was implemented using C++ and the OpenCV

library, which has an implementation of EMD. For su-

perpixel segmentation, the publicly available imple-

mentation given by Felzenszwalb et al(Felzenszwalb

and Huttenlocher, 2004) was used. On average, the

saliency computation took about 0.4 seconds on a sin-

gle core, 3 GHz Intel i5 processor with 8 GB RAM.

Data Sets: For the experiments, the standard

data sets ASD (Achanta et al., 2009), LSD (Li

et al., 2013), 2-SED (Alpert et al., 2007) and

complex-scene (CSD) (Yan et al., 2013) were used.

ASD and CSD are two of the largest data sets

publicly available with 1,000 images each. Each

image in ASD has a single stand out object while

each image in the CSD set has multiple stand

Algorithm 2: EMD-based saliency detection.

Input: Input Image

Output: Saliency Map

RGB → CIEL*a*b* ;

numRes → 20 ;

for j ← 0; j ≤ numRes do

Res[ j] → goodSegmentS cr() ;

BestSeg = max(Res[ j] → goodSegmentScr()) ;

N = BestSeg.size() ;

while i ≤ N do

Compute DCD

;

Compute sp

avgX,avgY;

Compute Image signature ;

Compute neighList

;

while i ≤ N do

sumSal = 0 ;

while k ≤ N do

Compute cost matrix ;

Compute emd

= ;

sumSal = sumS al + emdik;

Aggregate total difference ;

δ = ||sp

→ avgX,avgY − img →

avgX,avgY ||

;

Account for center bias Sal

(1 − δ) ;

out objects; each image in 2-SED has two salient

objects and LSD has images with different categories.

Results: The results of testing the proposed method

are shown in Fig. 5, in which each row shows the

results of a dataset, organized as triplets of an in-

put image, the result saliency map, and the ground

truth. The saliency map should be visually closer to

the ground truth image.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

562

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

FPR

TPR

ROC Curve

EMD

avg

clr

(a)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

FPR

TPR

ROC Curve

EMD

(b)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

FPR

TPR

ROC Curve

EMD

MRSD

(c)

Figure 6: ROC curves of comparing the proposed method (“EMD”) with (a) the baseline average color method, (b) other local

contrast methods, and (c) other global contrast methods.

3.2 Evaluation

We assess the performance of different methods as

follows. For each input image, we obtain the saliency

map from a given method and compare it to the

ground truth. The ground truth map is either on (“1”)

or off (“0”). Over a threshold value, when a ground

truth “1”-pixel is labeled as “salient,” that pixel is

considered a true positive. On the other hand, when a

ground truth “0”-pixel is labeled as “salient,” it is con-

sidered as false positive. We average the performance

over all images in a dataset. We plot the true positive

rate against the false positive rate to obtain the ROC

curve, which is used for quantitative evaluations. The

area under the curve shows how well the saliency

algorithm predicts against the ground truth, which is

the human ﬁxation data or the salient segment. The

following two sets of quantitative evaluation are done.

Baseline Comparison: It shows that EMD is a

better metric by comparing it to commonly used av-

erage color comparison. For consistency, the average

color map is normalized with the objectness map.

From Fig. 6a, it can be seen that the EMD-based

metric works better.

Comparison to State-of-the-Art Methods:

A comparison with local contrast-based saliency

detection methods CA (Goferman et al., 2010), IT

(Itti et al., 1998), GB (Harel et al., 2006), MZ (Ma

and Zhang, 2003), LC (Zhai and Shah, 2006), and

global contrast-based saliency detection methods AC

(Achanta et al., 2009), MRSD (Singh et al., 2014),

SF (Perazzi et al., 2012), RC (Cheng et al., 2011) on

the data set ASD (Achanta et al., 2009) is done. It

is worth noting that each of these methods is based

on contrasting the color information, either locally

or globally, to detect the salient regions. For this

data set, each author has publicly available results.

The ROC curve implementation given by (Borji and

Itti, 2013) is used to compare the proposed method

(“EMD”) to local methods and global methods,

shown in Fig. 6b and Fig. 6c, respectively.

Visual Comparison: A visual comparison of the

results is also presented to illustrate the performance

of the proposed method (“EMD”) compared to other

local contrast methods. (Fig. 7). The visual compari-

son shows that our method robustly highlights salient

segments for various input images.

4 CONCLUDING REMARKS

A novel saliency detection algorithm by ﬁnding the

amount of work done by moving dominant color de-

scriptors from the neighboring superpixel segments is

presented. This work overcame earlier problems of

edge highlighting associated with local comparison-

based methods. The EMD along with DCD captures

the color difference better hence resulting in a bet-

ter saliency map. Experimental results show that our

results are more consistent with the human vision’s

attention.

Our work shows a new promising direction of us-

ing EMD metric for local comparison-based saliency

detection with quantitative results better than local

comparison-based state-of-the-art methods. Ongoing

work includes comparing the information from the

objectness measure (Alexe et al., 2012) with other

methods of extracting objects recently proposed (Li

et al., 2014). Other ongoing work includes extending

the method to compute saliency not just from a single

frame but from a larger collection of images.

Comparing Color Descriptors between Image Segments for Saliency Detection

563

Input Image CA GB IT LC MZ EMD Ground Truth

Figure 7: Visual comparison with other state-of-the-art-methods. In each row, from left to right are an input image, followed

by results from methods CA (Goferman et al., 2010), GB (Harel et al., 2006), IT (Itti et al., 1998), LC (Cheng et al., 2011),

MZ (Ma and Zhang, 2003), result from the proposed EMD-based saliency method (“EMD”), and the ground truth.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

564

REFERENCES

Achanta, R., Hemami, S., Estrada, F., and S

usstrunk, S.

(2009). Frequency-tuned salient region detection. In

IEEE Conference on Computer Vision and Pattern

Recognition, pages 1597–1604.

Alexe, B., Deselaers, T., and Ferrari, V. (2012). Measur-

ing the objectness of image windows. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

34(11):2189–2202.

Alpert, S., Galun, M., Basri, R., and Brandt, A. (2007). Im-

age segmentation by probabilistic bottom-up aggrega-

tion and cue integration. In IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 1–8.

Bagon, S., Boiman, O., and Irani, M. (2008). What is a

good image segment? a uniﬁed approach to segment

extraction. In Proceedings of the 10th European Con-

ference on Computer Vision: Part IV, pages 30–44.

Ben-Av, M., Sagi, D., and Braun, J. (1992). Visual atten-

tion and perceptual grouping. Perception and Psy-

chophysics, 52(3):277–294.

Borji, A. and Itti, L. (2013). State-of-the-art in visual atten-

tion modeling. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 35(1):185–207.

Cheng, M.-M., Zhang, G.-X., Mitra, N. J., Huang, X., and

Hu, S.-M. (2011). Global contrast based salient region

detection. In IEEE Conference on Computer Vision

and Pattern Recognition, CVPR ’11, pages 409–416.

Engel, S., Zhang, X., and Wandell, B. (1997). Colour tun-

ing in human visual cortex measured with functional

magnetic resonance imaging. Nature, 388(6637):68–

71.

Felzenszwalb, P. and Huttenlocher, D. (2004). Efﬁcient

graph-based image segmentation. International Jour-

nal Computer Vision, 59(2):167–181.

Goferman, S., Zelnik-Manor, L., and Tal, A. (2010).

Context-aware saliency detection. In IEEE Confer-

ence on Computer Vision and Pattern Recognition,

pages 2376–2383.

Harel, J., Koch, C., and Perona, P. (2006). Graph-based

visual saliency. In Advances in Neural Information

Processing Systems, pages 545–552. MIT Press.

Itti, L., Koch, C., and Niebur, E. (1998). A model of

saliency-based visual attention for rapid scene anal-

ysis. IEEE Transactions on Pattern Analysis and Ma-

chine Intelligence, 20(11):1254–1259.

Li, J., Levine, M., An, X., Xu, X., and He, H. (2013). Vi-

sual saliency based on scale-space analysis in the fre-

quency domain. IEEE Transactions on Pattern Anal-

ysis and Machine Intelligence, 35(4):996–1010.

Li, Y., Hou, X., Koch, C., Rehg, J., and Yuille, A. (2014).

The secrets of salient object segmentation. In IEEE

Conference on Computer Vision and Pattern Recogni-

tion, pages 4321–4328.

Lin, Y., Tang, Y., Fang, B., Shang, Z., Huang, Y., and

Wang, S. (2013). A visual-attention model using

earth mover’s distance-based saliency measurement

and nonlinear feature combination. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

35(2):314–328.

Liu, T., Sun, J., Zheng, N.-N., Tang, X., and Shum, H.-Y.

(2007). Learning to detect a salient object. In IEEE

Conference on Computer Vision and Pattern Recogni-

tion, pages 1–8.

Ma, Y.-F. and Zhang, H.-J. (2003). Contrast-based image

attention analysis by using fuzzy growing. In ACM

International Conference on Multimedia, pages 374–

381.

Manjunath, B. S., Ohm, J.-R., Vasudevan, V. V., and Ya-

mada, A. (2001). Color and texture descriptors. IEEE

Transactions on Circuits and Systems for Video Tech-

nology, 11(6):703–715.

Perazzi, F., Krahenbuhl, P., Pritch, Y., and Hornung, A.

(2012). Saliency ﬁlters: Contrast based ﬁltering for

salient region detection. In IEEE Conference on Com-

puter Vision and Pattern Recognition, pages 733–740.

Rubner, Y., Tomasi, C., and Guibas, L. (2000a). The earth

mover’s distance as a metric for image retrieval. Inter-

national Journal of Computer Vision, 40(2):99–121.

Rubner, Y., Tomasi, C., and Guibas, L. J. (2000b). The

earth mover’s distance as a metric for image retrieval.

International Journal of Computer Vision, 40(2):99–

121.

Singh, A., Chu, C., and Pratt, M. A. (2014). Multiresolu-

tion superpixels for visual saliency detection. In IEEE

Symposium on Computational Intelligence for Multi-

media, Signal and Vision Processing, pages 1–8.

Sun, J. and Ling, H. (2013). Scale and object aware im-

age thumbnailing. International Journal of Computer

Vision, 104(2):135–153.

Treisman, A. (1982). Perceptual grouping and attention in

visual search for features and for objects. Journal

of Experimental Psychology: Human Perception and

Performance, 8(2):194.

Yan, Q., Xu, L., Shi, J., and Jia, J. (2013). Hierarchical

saliency detection. In IEEE Conference on Computer

Vision and Pattern Recognition, pages 1155–1162.

Yang, N.-C., Chang, W.-H., Kuo, C.-M., and Li, T.-H.

(2008). A fast mpeg-7 dominant color extraction with

new similarity measure for image retrieval. Journal

of Visual Communication and Image Representation,

19(2):92–105.

Zhai, Y. and Shah, M. (2006). Visual attention detection in

video sequences using spatiotemporal cues. In ACM

International Conference on Multimedia, pages 815–

824.

Comparing Color Descriptors between Image Segments for Saliency Detection

565