Hindawi Publishing Corporation

Computational Intelligence and Neuroscience

Volume 2015, Article ID 875735, 8 pages
http://dx.doi.org/10.1155/2015/875735

Research Article
Saliency Mapping Enhanced by Structure Tensor

Zhiyong He,1 Xin Chen,2 and Lining Sun1

1
School of Mechanical and Electric Engineering, Soochow University, Suzhou 215021, China
2
NovuMind Inc., Santa Clara, CA 95131, USA

Correspondence should be addressed to Zhiyong He; he-zhiyong@139.com

Received 4 June 2015; Revised 11 September 2015; Accepted 27 September 2015

Academic Editor: Paolo Del Giudice

Copyright © 2015 Zhiyong He et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We propose a novel efficient algorithm for computing visual saliency, which is based on the computation architecture of Itti
model. As one of well-known bottom-up visual saliency models, Itti method evaluates three low-level features, color, intensity,
and orientation, and then generates multiscale activation maps. Finally, a saliency map is aggregated with multiscale fusion. In
our method, the orientation feature is replaced by edge and corner features extracted by a linear structure tensor. Following it,
these features are used to generate contour activation map, and then all activation maps are directly combined into a saliency map.
Compared to Itti method, our method is more computationally efficient because structure tensor is more computationally efficient
than Gabor filter that is used to compute the orientation feature and our aggregation is a direct method instead of the multiscale
operator. Experiments on Bruce’s dataset show that our method is a strong contender for the state of the art.

1. Introduction of forty activation maps. Finally, a linear operator is employed

to normalize these maps, in which the most salient location is
Visual saliency (also called visual salience) refers to the quality selected by the winner-take-all neural network to generate a
or state by which information stands out relative to its neigh- saliency map. However, results of Itti method are sometimes
bors and often attracts human attention [1]. In the subsequent blurry and prefer small and local features, which are less
stages, the salient images are preferentially taken as inputs useful for some further computer vision applications such as
instead of the whole image. As a consequence, visual saliency object segmentation and detection.
has been widely applied to various computer vision tasks such Despite many advances of the visual saliency made in
as segmentation [2, 3], image retargeting [4–6], object detec- recent years, the various evaluation results in [14] indicate
tion [7, 8], image collection [9], and object recognition [10]. that there are still some questions about the mechanism of
Koch and Ullman introduced a basic biologically inspired visual saliency. In addition to a motivation of investigation
architecture of visual saliency, referred to as Koch and Ullman on some questions such as low-level features and the combi-
model [11]. Then Itti et al. presented a computational archi- nation of activation maps, in this paper, we also focus on the
tecture to implement and verify the Koch and Ullman model performance of algorithm and whether the saliency results
[12]. As summarized in [13], most of the implementation can greatly benefit computer vision applications.
techniques of the visual saliency models generally have three Results from some recent research works have shown that
stages: (1) extraction: extracting low-level features at locations features of edge and corner also play important roles in visual
over the image plane, (2) activation: forming activation saliency [13, 15, 16]. In our study, we also note that orientation
maps from the features, and (3) normalization/combination: features are less likely to win in the combination of activation
normalizing the activation maps and combining them into a maps. Moreover, Gabor filters used for orientation extraction
single saliency map. are computationally expensive.
For Itti method, the objective of the first stage is to extract We therefore propose an efficient method to compute
three low-level features, intensity, color, and orientation, and saliency map, referred to as structure tensor (ST) saliency
followed by using Difference of Gaussian (DOG) to form total map. The computational architecture of our method is shown
2 Computational Intelligence and Neuroscience

Input image

Linear filtering Structure tensor

Colors Intensity Edge Corner

Center-surround differences and Normalizing the values to a

normalization fixed range

Feature Maps

Across-scale combinations Linear combination

Activation Maps
Colors Intensity Contour

Normalization and linear

combination

Saliency map

Figure 1: General architecture of our method. In our method, we call the activation map generated by edge and corner features contour
activation map. The final saliency map combined these activation maps into ST saliency map.

in Figure 1 and is the same as Itti method in feature In Section 4, we present our experimental results and quanti-
extraction and activation maps generation for intensity and tative evaluations on a challenging dataset and discuss them.
color features. The features of edge and corner are extracted This paper closes with a conclusion of our work in Section 5.
by structure tensor and directly combined into an activation
map, called contour map. After obtaining three activation 2. Related Work
maps, we use linear combination to aggregate activation maps
to a saliency map instead of multiscale combination and Visual saliency methods are generally categorized into bio-
winner-take-all rule. logically inspired methods and computationally oriented
This paper makes two major contributions as follows: methods. There is an extensive literature on the areas, but
here we mention just a few relevant papers. Some surveys are
(1) We propose a novel efficient algorithm to calculate the found in [17–19], and some recent progress is reported in [20].
saliency map. Compared to other methods performed Koch and Ullman [11] proposed a basic architecture of
on a challenging dataset, besides the best performance biologically inspired methods and defined a saliency map
achieved, the results of our method obtain sharper as a topographic map that represents conspicuousness of
boundaries which are useful in some further applica- scene locations. Their work also introduced a winner-take-
tions such as object segmentation and detection. all neural network that selects the most salient location and
(2) Our work has shown that edge and corner are two employs an inhibition of return mechanism to allow the focus
important low-level features in saliency generation. of attention to shift to the next most salient location. Then
Itti et al. presented a computational model to implement
The paper is organized as follows. Section 2 briefly reviews and verify Koch and Ullman model. Since then, the works
the state-of-the-art methods with particular emphasis on related to the saliency map have quickly become one of the
saliency algorithms related to Itti method, and Section 3 hot research fields.
introduces some backgrounds of structure tensor and for- Itti method employs a Difference of Gaussian (DOG)
mally describes our algorithm of saliency map computation. operator to evaluate color, intensity, and orientation features
Computational Intelligence and Neuroscience 3

to generate total of forty activation maps and across-scale-

combines these maps into a saliency map. Besides the
expensive computation, one big problem of Itti method is that
the results are sometimes blurry and prefers small purely local E1
features. On the other hand, many algorithms of computer
vision need input features related to contours because they
require the distinct boundary information. Recently, some
methods have been proposed to obtain sharp edges, for exam- E2
ple, local dissimilarities at the pixel level [21], multiscale DOG
[22], and histogram analysis [15]. However, the common
problem of these methods is that they are more sensitive to
the noises.
As mentioned in the previous section, improving on Itti Figure 2: The relation between ellipse and structure tensor.
method, we propose an efficient algorithm for calculating
the saliency map, and the computational architecture of our
method is shown in Figure 1. The computational architecture
of our method is similar to Itti method, and our method = 𝐼 (𝑢, V)
evaluates intensity, color, edge, and corner features instead of
intensity, color, and orientation features. The structure tensor 𝜕𝐼 𝜕𝐼 Δ𝑥
+[ (𝑢, V) , (𝑢, V)] [ ] .
is used to extract the features of the edge and corner. In the 𝜕𝑥 𝜕𝑦 Δ𝑦
final step, we use the linear combination to generate a saliency
(2)
map instead of the winner-take-all rule of Itti method.

Hence, (1) can be rewritten as

3. The Proposed Saliency Model
2
In this section, we briefly introduce the background of 𝐸 = ∑ (𝐼 (𝑢, V) − 𝐼 (𝑢 + Δ𝑥, V + Δ𝑦))
structure tensor and formally describe our algorithm. 𝑤
(3)
Δ𝑥
3.1. Introduction to Structure Tensor. In mathematics, struc- ≈ [Δ𝑥, Δ𝑦] T [ ],
Δ𝑦
ture tensor is a matrix representation of partial derivative
information. In the field of image processing and computer
vision, it typically represents the gradient or edge information where matrix T is
and has a more powerful description of local patterns as
opposed to the directional derivative through its coherence 𝜕𝐼 2 𝜕𝐼 𝜕𝐼
measure [23, 24]. [ 𝜕𝑥 ) 𝜕𝑥 ⋅ 𝜕𝑦 ]
(
[ ]
There are two categories of structure tensor: linear struc- T = ∑[[
]
]
ture tensor and nonlinear structure tensor. Compared to the 𝑤 [ 𝜕𝐼 𝜕𝐼 𝜕𝐼 2 ]
nonlinear structure tensor, the linear structure tensor is fast ⋅ ( )
[ 𝜕𝑥 𝜕𝑦 𝜕𝑦 ]
and easy to implement with Fast Fourier Transform (FFT). (4)
We therefore select the linear structure tensor to extract the 𝜕𝐼 2 𝜕𝐼 𝜕𝐼
features of edges and corners. [ ∑ ( 𝜕𝑥 ) ∑ 𝜕𝑥 ⋅ 𝜕𝑦 ]
[𝑤 𝑤 ]
Given an image 𝐼(𝑥, 𝑦), if pixel (𝑥, 𝑦) translates to (𝑥 + =[
[
].
]
[ 𝜕𝐼 𝜕𝐼 2
Δ𝑥, 𝑦 + Δ𝑦), the energy 𝐸 is defined as 𝜕𝐼 ]
∑ ⋅ ∑( )
[ 𝑤 𝜕𝑥 𝜕𝑦 𝑤 𝜕𝑦 ]
2
𝐸= ∑ 𝑤 (𝑢, V) (𝐼 (𝑢, V) − 𝐼 (𝑢 + Δ𝑥, V + Δ𝑦)) , (1)
(𝑢,V)∈𝑊(𝑥,𝑦) Matrix T is a structure tensor, which is also considered as a
covariance matrix.
where 𝑊(𝑥, 𝑦) is a window at center point (𝑥, 𝑦) and 𝑤(𝑢, V) We also consider (3) as an approximation of a binomial
is a weight function at pixel (𝑥, 𝑦). In the rest of this paper, function, and from a view of geometry, a binomial function
∑(𝑢,V)∈𝑊(𝑥,𝑦) 𝑤(𝑢, V) is simply written as ∑𝑤 . is an ellipse where short axis and long axis are represented as
eigenvalues 𝐸1 and 𝐸2, respectively. The direction of ellipse
It is approximated by a first-order Taylor series: is determined by the eigenvectors. As shown in Figure 2, the
𝜕𝐼 equation of ellipse is written as
𝐼 (𝑢 + Δ𝑥, V + Δ𝑦) ≈ 𝐼 (𝑢, V) + (𝑢, V) Δ𝑥
𝜕𝑥
𝜕𝐼 Δ𝑥
+ (𝑢, V) Δ𝑦 [Δ𝑥, Δ𝑦] T [ ] = 1. (5)
𝜕𝑦 Δ𝑦
4 Computational Intelligence and Neuroscience

(1) Input:
(2) Input image I: three-channel and size (𝑚, 𝑛)
(3) Output:
(4) Edge feature map A: one channel and size (𝑚1, 𝑛1)
(5) Corner feature map B: one channel and size (𝑚1, 𝑛1)
(6) Contour activation map C: one channel and size (𝑚1, 𝑛1)
(7) Begin
(8) Resize the input image I to (𝑚1, 𝑛1), called Im-Re
(9) for 𝑗 ← 1, 𝑛1 do
(10) for 𝑖 ← 1, 𝑚1 do
(11) For Im-Re, calculate structure tensor 𝐽𝜎 using (6)
(12) Calculate eigenvalues 𝜆 1 and 𝜆 2 using (7) and (8), respectively
(13) 𝐴(𝑖, 𝑗) = 𝜆 1 − 𝜆 2
(14) 𝐵(𝑖, 𝑗) = 𝜆 1 + 𝜆 2
(15) end for
(16) end for
(17) Normalize A and B into a fixed range [0⋅ ⋅ ⋅ 1]
(18) Combine normalized A and normalized B into CT
(19) End

Algorithm 1: Algorithm of contour activation map.

Based on (4), some types of structure tensor have been In the final step, we combine feature maps into a contour
constructed. In our work, we use a linear structure tensor to activation map CT as follows:
analyze the input image, and it is defined as
1
CT = (𝑁 (𝐴) + 𝑁 (𝐵)) , (10)
𝜕𝐼 2
𝜕𝐼 𝜕𝐼 2
3 [
𝐾𝜎 ∗ ( 𝑖 ) 𝐾𝜎 ∗ ( 𝑖 ⋅ 𝑖 )]
[ 𝜕𝑥 𝜕𝑥 𝜕𝑦 ] where 𝑁(𝐴) is the normalized edge feature map and 𝑁(𝐵) is
𝐽𝜎 = ∑ [ [
],
] (6) the normalized corner feature map.
𝑖=1 [ 𝜕𝐼𝑖 𝜕𝐼𝑖 𝜕𝐼𝑖 2 ]
𝐾𝜎 ∗ ( ⋅ ) 𝐾𝜎 ∗ ( )
[ 𝜕𝑥 𝜕𝑦 𝜕𝑦 ] 3.3. ST Saliency Map Generation. We assume that all features
equally contribute to the ST saliency map generation. After
where 𝐾𝜎 is a Gaussian kernel with variance 𝜎 and ∗ is a obtaining the contour activation map, the intensity activation
convolution operator. The parameter 𝑖 is the image channel map, and the color activation map, we combine them into a
number. saliency map as follows:
For any kind of structure tensor, we use [ 𝐺𝐹 𝐻
𝐹 ] to simply

represent matrix T of (4). Then the two eigenvalues are 1

𝑆= (⨁ (CL) + ⨁ (𝐼) + ⨁ (CT)) , (11)
calculated as 3
where ⨁ is a normalization operation which is defined
𝐺 + 𝐻 + √(𝐺 − 𝐻)2 + 4𝐹2 (7) in [12], CL is the color activation map, I is the intensity
𝜆1 = ,
2 activation map, and CT is the contour activation map.
Some saliency maps of our method are shown in Figure 3
𝐺 + 𝐻 − √(𝐺 − 𝐻)2 + 4𝐹2 (8)
and these maps have distinct boundaries.
𝜆2 = .
2
4. Experimental Results
3.2. Contour Activation Map. As shown in Figure 1, we
calculate the activation maps of color and intensity with In this section, we present subjective evaluation and quan-
Itti method, and the contour activation map is detailed in titative analysis of our method and some state-of-the-art
Algorithm 1. methods and analysis of performance of these methods.
For computation of A and B, we do not need to compute 4.1. Saliency Maps. We compared saliency maps of our
𝜆 1 and 𝜆 2 with (7) and (8) and add and subtract these values method with saliency maps of some state-of-the-art methods
to calculate A and B. We directly compute them as including Itti method [12], Attention Based on Information
Maximization (AIM) method [25, 26], Dynamic Visual Atten-
𝜆 1 − 𝜆 2 = √(𝐺 − 𝐻)2 + 4𝐹2 , tion (DVA) method [27], Graphic-Based Visual Saliency
(9) (GBVS) method [13], and Image Signature (IS) method [28].
𝜆 1 + 𝜆 2 = 𝐺 + 𝐻. The MATLAB implementation of these methods is based on
Computational Intelligence and Neuroscience 5

Figure 3: Saliency maps of our method. The odd row is the input images from Bruce dataset, and the even row is the saliency generated by
our method. Obviously, the saliency maps of our method have sharp edges.
6 Computational Intelligence and Neuroscience

(a) (b) (c) (d) (e) (f) (g)

Figure 4: Saliency maps on the Bruce dataset. (a) Input image, (b) our method, (c) Itti method, (d) AIM method, (e) DVA method, (f) GBVS
method, and (g) IS method using LAB color space. Since our method includes the edge and corner information, saliency maps of our method
have sharp edges that are useful for the further steps in some computer vision tasks.

the codes on the authors’ websites. Saliency maps are shown remove this center bias, following the procedure of Tatler et
in Figure 4. al.’s work [29], Hou et al. [28] introduced ROC Area Under
the Curve (AUC) score to quantitatively evaluate the perfor-
4.2. Analysis of Performance. We evaluated our method on mance of different algorithms. Good results should maximize
Bruce dataset containing 120 natural images with eye fixation the ROC AUC score. To compare the ROC AUC scores, we
ground truth data. In Bruce dataset, the size of all images is follow the computation method provided by [28], but the size
681 × 511. Some of methods are sensitive to different sizes of (170 × 128) is different with two input image sizes used in [28].
the input image. As a consequence, in order to fairly evaluate Comparison of the ROC AUC scores is shown in Figure 5.
results of different methods, we resize the input images to the We conducted our tests on a laptop with Intel Dual-Core
same size (170 × 128) for each method. i5-4210U 1.7 GHz CPU and 4 G RAM memory. All codes were
Results from perceptual research works [29, 30] have written in MATLAB.
found that human fixations have strong center bias which The execution times of the methods are summarized in
may affect the performance of a saliency algorithm. To Figure 6, in which the time is an average time of 120 images.
Computational Intelligence and Neuroscience 7

ROC score dependence on blur tensor to extract these features. The reason that our algorithm
0.75
is efficient lies in the following: (1) linear structure tensor is
(averaged across 120 images)

an efficient feature extraction algorithm and (2) our linear

0.7
combination method is fast. On the basis of experimental
ROC score

results on Bruce dataset, our method has shown that some

0.65
computer vision tasks, in particular real time applications,
can benefit from our method as a preprocessing step.
0.6

0.55 Conflict of Interests

0.02 0.04 0.06 0.08 0.1 0.12
Blur width The authors declare that there is no conflict of interests
(STD of Gaussian kernel in image widths) regarding the publication of this paper.
Proposed DVA
Itti GBVS Acknowledgments
AIM
Figure 5: Comparison of the ROC AUC scores of all methods. Our This work was supported in part by the Chinese National Nat-
method achieves the best ROC AUC score. ural Science Foundation (NSFC 61473201, 51405320), the Nat-
ural Science Foundation of Jiangsu Province (BK20150339),
and the Science and Technology Program of Suzhou
Computational complexity comparison (SYG201424).
Average computation time (s)

2.415 References
1 [1] M. Carrasco, “Visual attention: the past 25 years,” Vision
0.842 Research, vol. 51, no. 13, pp. 1484–1525, 2011.
0.8
[2] J. Han, K. N. Ngan, M. Li, and H.-J. Zhang, “Unsupervised
0.6
0.406
extraction of visual attention objects in color images,” IEEE
0.4 0.353 0.318 Transactions on Circuits and Systems for Video Technology, vol.
0.2 0.199 16, no. 1, pp. 141–145, 2006.
0 [3] E. Rahtu, J. Kannala, M. Salo, and J. Heikkilä, “Segmenting
salient objects from images and videos,” in Computer Vision—
Proposed DVA ECCV 2010, vol. 6315 of Lecture Notes in Computer Science, pp.
Itti GBVS 366–379, Springer, Berlin, Germany, 2010.
AIM Itti-orientation [4] S. Avidan and A. Shamir, “Seam carving for content-aware
Figure 6: Results of the performance of these different methods. image resizing,” ACM Transactions on Graphics, vol. 26, no. 3,
Time measurements are given in seconds. The results are the average article 10, 2007.
times of 120 images of Bruce dataset. [5] S. Goferman, L. Zelnik-Manor, and A. Tal, “Context-aware
saliency detection,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 34, no. 10, pp. 1915–1926, 2012.
[6] D. Vaquero, M. Turk, K. Pulli, M. Tico, and N. Gelfand, “A
The figure shows that our method is about twice as fast as survey of image retargeting techniques,” in Applications of
Itti method and outperforms other state-of-the-art methods. Digital Image Processing XXXIII, vol. 7798 of Proceedings of
The reason lies in two parts. First, structure tensor is an SPIE, pp. 779–814, SPIE Optical Engineering + Applications,
efficient algorithm of feature extraction. Second, we directly San Diego, Calif, USA, August 2010.
aggregate three activation maps into a saliency map. It is [7] A. Oliva, A. Torralba, M. S. Castelhano, and J. M. Henderson,
obvious that the performance will increase greatly if our “Top-down control of visual attention in object detection,” in
method is implemented by C/C++, and it should satisfy most Proceedings of the International Conference on Image Processing
of the real time applications. (ICIP ’03), pp. I-253–I-256, September 2003.
[8] X. Shen and Y. Wu, “A unified approach to salient object
detection via low rank matrix recovery,” in Proceedings of the
5. Conclusion IEEE Conference on Computer Vision and Pattern Recognition
(CVPR ’12), pp. 853–860, Providence, RI, USA, June 2012.
In this paper we have proposed an efficient algorithm for [9] M.-M. Cheng, N. J. Mitra, X. Huang, and S.-M. Hu,
computing the saliency map, which has a distinct boundary “SalientShape: group saliency in image collections,” The
that contributes to further computer vision applications such Visual Computer, vol. 30, no. 4, pp. 443–453, 2014.
as segmentation and detection. The computational architec- [10] U. Rutishauser, D. Walther, C. Koch, and P. Perona, “Is bottom-
ture of our method is close to Itti method, but we have made up attention useful for object recognition?” in Proceedings of
two improvements in low-level features extraction and com- the IEEE Computer Society Conference on Computer Vision and
bination of activation maps. Since features of edge and corner Pattern Recognition (CVPR ’04), vol. 2, pp. II-37–II-44, IEEE,
are important cues in visual saliency, we use a linear structure July 2004.
8 Computational Intelligence and Neuroscience

[11] C. Koch and S. Ullman, “Shifts in selective visual attention: [28] X. Hou, J. Harel, and C. Koch, “Image signature: highlighting
towards the underlying neural circuitry,” Human Neurobiology, sparse salient regions,” IEEE Transactions on Pattern Analysis
vol. 4, no. 4, pp. 219–227, 1985. and Machine Intelligence, vol. 34, no. 1, pp. 194–201, 2012.
[12] L. Itti, C. Koch, and E. Niebur, “A model of saliency-based visual [29] B. W. Tatler, R. J. Baddeley, and I. D. Gilchrist, “Visual correlates
attention for rapid scene analysis,” IEEE Transactions on Pattern of fixation selection: effects of scale and time,” Vision Research,
Analysis and Machine Intelligence, vol. 20, no. 11, pp. 1254–1259, vol. 45, no. 5, pp. 643–659, 2005.
1998. [30] L. Zhang, M. H. Tong, T. K. Marks, H. Shan, and G. W.
[13] J. Harel, C. Koch, and P. Perona, “Graph-based visual saliency,” Cottrell, “SUN: a bayesian framework for saliency using natural
in Proceedings of the Advances in Neural Information Processing statistics,” Journal of Vision, vol. 8, no. 7, article 32, 2008.
Systems (NIPS ’06), pp. 545–552, Vancouver, Canada, December
2006.
[14] M.-M. Cheng, G.-X. Zhang, N. J. Mitra, X. Huang, and S.-
M. Hu, “Global contrast based salient region detection,” in
Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR ’11), pp. 409–416, Providence, RI,
USA, June 2011.
[15] T. Liu, Z. Yuan, J. Sun et al., “Learning to detect a salient object,”
IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 33, no. 2, pp. 353–367, 2011.
[16] R. Valenti, N. Sebe, and T. Gevers, “Image saliency by isocentric
curvedness and color,” in Proceedings of the 12th IEEE Inter-
national Conference on Computer Vision, pp. 2185–2192, IEEE,
Kyoto, Japan, September-October 2009.
[17] A. Borji and L. Itti, “State-of-the-art in visual attention mod-
eling,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 35, no. 1, pp. 185–207, 2013.
[18] A. Borji, H. R. Tavakoli, D. N. Sihite, and L. Itti, “Analysis
of scores, datasets, and models in visual saliency prediction,”
in Proceedings of the 14th IEEE International Conference on
Computer Vision (ICCV ’13), pp. 921–928, Sydney, Australia,
December 2013.
[19] S. Frintrop, E. Rome, and H. I. Christensen, “Computational
visual attention systems and their cognitive foundations: a
survey,” ACM Transactions on Applied Perception, vol. 7, no. 1,
article 6, 2010.
[20] Z. Bylinskii, T. Judd, A. Borji et al., “MIT Saliency Benchmark,”
2015, http://saliency.mit.edu/index.html.
[21] Y.-F. Ma and H.-J. Zhang, “Contrast-based image attention
analysis by using fuzzy growing,” in Proceedings of the 11th ACM
International Conference on Multimedia (MM ’03), pp. 374–381,
ACM, November 2003.
[22] L. Itti and P. F. Baldi, “Bayesian surprise attracts human atten-
tion,” in Advances in Neural Information Processing Systems, pp.
547–554, MIT Press, 2005.
[23] T. Brox, J. Weickert, B. Burgeth, and P. Mrázek, “Nonlinear
structure tensors,” Image and Vision Computing, vol. 24, no. 1,
pp. 41–55, 2006.
[24] U. Köthe, “Edge and junction detection with an improved
structure tensor,” in Pattern Recognition, pp. 25–32, Springer,
Berlin, Germany, 2003.
[25] N. Bruce and J. Tsotsos, “Saliency based on information maxi-
mization,” in Proceedings of the Advances in Neural Information
Processing Systems (NIPS ’05), pp. 155–162, Vancouver, Canada,
December 2005.
[26] N. D. B. Bruce and J. K. Tsotsos, “Saliency, attention and visual
search: an information theoretic approach,” Journal of Vision,
vol. 9, no. 3, article 5, 2009.
[27] X. Hou and L. Zhang, “Dynamic visual attention: searching for
coding length increments,” in Advances in Neural Information
Processing Systems, pp. 681–688, MIT Press, 2009.
 Advances in Journal of
Industrial Engineering
Multimedia
Applied
Computational
Intelligence and Soft
Computing
The Scientific International Journal of
Distributed
Hindawi Publishing Corporation
World Journal
Hindawi Publishing Corporation
Sensor Networks
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Advances in

Fuzzy
Systems
Modelling &
Simulation
in Engineering
Hindawi Publishing Corporation
Hindawi Publishing Corporation Volume 2014 http://www.hindawi.com Volume 2014
http://www.hindawi.com

Submit your manuscripts at

Journal of
http://www.hindawi.com
Computer Networks
and Communications Advances in
Artificial
Intelligence
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014

International Journal of Advances in

Biomedical Imaging Artificial
Neural Systems

International Journal of
Advances in Computer Games Advances in
Computer Engineering Technology Software Engineering
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

International Journal of
Reconfigurable
Computing

Advances in Computational Journal of

Journal of Human-Computer Intelligence and Electrical and Computer
Robotics
Hindawi Publishing Corporation
Interaction
Hindawi Publishing Corporation
Neuroscience
Hindawi Publishing Corporation Hindawi Publishing Corporation
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014