Advertisement

Vision Saliency Feature Extraction Based on Multi-scale Tensor Region Covariance

  • Shimin WangEmail author
  • Mingwen Wang
  • Jihua Ye
  • Anquan Jie
Conference paper
  • 400 Downloads
Part of the Lecture Notes in Computer Science book series (LNCS, volume 10390)

Abstract

In the process of extracting image saliency features by using regional covariance, the low-level higher-order data are dealt with by vectorization, however, the structure of the data (color, intensity, direction) may be lost in the process, leading to a poorer representation and overall performance degradation. In this paper we introduce an approach for sparse representation of region covariance that will preserve the inherent structure of the image. This approach firstly calculates the image low-level data (color, intensity, direction), and then uses multi-scale transform to extract the multi-scale features for constructing tensor space, at last by using tensor sparse coding the image bottom features are extracted from region covariance. In the paper, it compares the experimental results with the commonly used feature extraction algorithms’ results. The experimental results show that the proposed algorithm is closer to the actual boundary of the object and achieving better results.

Keywords

Saliency feature Region covariance Tensor space Multi-scale transform 

1 Introduction

When people observe natural scenes, predictions of objects are called saliency predictions or detections, and the saliency studying has attracted a lot of computer vision scientists and neuroscience scientists. Highly predictive accuracy is important for many applications, such as object detection and segmentation, content-based compression, scene understanding, robot navigation, and image/video quality assessment.

In the existing literatures the saliency models include: (1) pure bottom-up image saliency features; (2) top-down semantic saliency features; (3) combination models. These models use a variety of visual features, including low, medium and advanced features. However, each single model has its own assumptions and methods, respectively, to solve the different aspects of human visual attention to the problem. In addition, some researchers selected fewer data sets to validate their models. Thus, there is a data set deviation in their model [1]. Because of these assumptions and constraints, each model has its own image set [2] and the corresponding weaknesses. The bottom-up saliency model of the use of biological low-level features is mainly based on the principle proposed by Itti and others [3, 4, 5]. Assuming that the saliency region is significant in terms of color, intensity, or orientation, Itti et al. used the visual saliency of the center-surround difference across multi-scale image features. Harel and others [6] based on the local image block different from the surrounding image block to establish the image of the visual saliency (model GBVS), which uses the Markov chain to measure the dissimilarity between local blocks. Similarly, Liu and others [7] proposed a set of novel features including multi-scale contrast, center-surround histogram, and color spatial distribution to describe a salient object locally, regionally, and globally. Bruce and Tsotsos [8] presented a saliency model (AIM) is based on Shannon’s self-information measure and is achieved in a neural circuit, which is demonstrated as having close ties with the circuitry existent in the primate visual cortex. Hou and Zhang [9] introduced the incremental coding length (ICL) to measure the perspective entropy gain of each feature. Zhang and others [10] proposed a definition of saliency by considering what the visual system is trying to optimize when directing attention. Kong and others [11] proposed an integration approach to detect salient region based on three principles from psychological evidence and observations of images, including colour contrast in a global context, spatially compact colour distribution, and multi-scale image abstraction. Fang and others [12] proposed a novel video saliency detection model based on feature contrast in compressed domain. Four types of features including luminance, color, texture, and motion are extracted from the discrete cosine transform coefficients and motion vectors in video bitstream. Jiang and others [13] proposed a novel automatic salient object segmentation algorithm which integrates both bottom-up salient stimuli and object-level shape prior. The above models perform well, and each model has its own most suitable image set due to different assumptions. For example, the Itti [3] model focuses primarily on local dissimilarity and is insufficient in detecting global visual saliency, whereas the AIM [8] model shows insufficient performance in detecting local dissimilarity.

In general, these bottom-up models cannot detect saliency areas with semantic information. Itti model cannot distinguish between the human body and its surroundings. A large number of researchers recently solved the saliency detection by combining low-level and advanced features. Cerf and others [14] improved the performance by adding face detection to the Itti model. Judd and others [15] proposed a new saliency detection model by using the support vector machine (SVM) to learn the optimal weight of all the combined features. Similarly, Chang and others [16, 17] proposed object-based saliency models using object metrics. While the top-down model emphasizes the importance of advanced and semantic features, the goal of detecting failure is usually a saliency object that has not yet been trained. In addition, the performance of the combined model in detecting non-semantic bottom-up saliency regions may not be as good as pure bottom-up model. For example, the Judd [15] model does not well detect local dissimilarity in some cases.

Before the study most of algorithms dealt with the low-level higher-order data by vectorization, the structure of the data may be lost in the process, which leads to a poorer representation and overall performance degradation. In this paper we use regional covariance to extract image saliency features, and we introduce an approach for sparse representation of region covariance which will preserve the inherent structure of the image. In the paper, it compares the experimental results with the commonly used feature extraction algorithms. The experimental results show that the proposed algorithm is closer to the actual boundary of the object and achieving better results.

2 Algorithm

In the process of saliency image feature extraction, it is necessary to save the structural features of the saliency region texture as much as possible. At the same time, the local low-level features of the image should be preserved. From the statistical point of view, the features of visual saliency detections are very similar to the statistical components with the Supper Gaussian distribution, because the “super-Gaussian” in statistics and “sparse, structured” in image processing basically have the same meaning. This approach uses multi-scale transform to extract the multi-scale features for constructing tensor space, at last by tensor sparse coding the image bottom features are extracted from region covariance.

2.1 Region Covariance

2006 Erkut and Aykut [18] first proposed covariance as a compact regional descriptor, and since then it has been effectively using and solving various advanced computer vision problems. Description of region covariance:
$$ \varvec{Q}(x,y) = \varvec{F}(\varvec{I},x,y) $$
(1)
The \( \varvec{I} \) represents an image, the target area in the image is \( R \), its size is \( M \times N \), each pixel to generate \( d \)-dimension eigenvector \( \varvec{Q}(x,y) \). The \( M \) and \( N \) is the width and height of the target area, \( \varvec{Q}(x,y) \) is a description of the pixel in the target area, and its element may be the spatial position information, the intensity, the color, the direction, and may be the response value of the filter of the pixel, etc. The choice of eigenvectors can be determined according to the needs of the actual application. A pixel \( \varvec{Q}_{i} \) can be expressed as follows:
$$ \varvec{Q}_{i} = [x,y,R(x,y),G(x,y),B(x,y),T(x,y), \ldots ] $$
(2)
The \( (x,y) \) is the coordinates of the pixel; \( R(x,y) \), \( G(x,y) \), \( B(x,y) \) are the corresponding pixel \( (x,y) \) RGB color values; \( T(x,y) \) is the corresponding pixel brightness feature. The covariance matrix of the target region can be expressed as:
$$ \varvec{C}_{R} = \frac{1}{MN}\sum\limits_{i = 1}^{MN} {(\varvec{Q}_{i} - \varvec{u}_{R} )^{T} } (\varvec{Q}_{i} - \varvec{u}_{R} ) $$
(3)

Formula: \( \varvec{u}_{R} = \frac{1}{MN}\sum\limits_{i = 1}^{MN} {\varvec{Q}_{i} } \), the covariance matrix \( \varvec{C}_{R} \) is a real symmetric positive definite matrix, the values on the diagonal in the covariance matrix represent each individual feature, and the value on the non-diagonal indicates the correlation between the features, The size of \( {\mathbf{C}}_{R} \) depends only on the dimension of the selected feature vector \( d \), regardless of the size of the target area.

2.2 Multi-scale of Image Low-Level Features

The input image is assumed to be \( \varvec{I}(x,y) \), the red, green, and blue channel images are (\( \varvec{R}_{0} \varvec{,G}_{0} \varvec{,B}_{0} \)).

Color features: set up \( \varvec{R} \) (red), \( \varvec{G} \) (green), \( \varvec{B} \) (blue), \( \varvec{Y} \) (yellow) 4 color channels, such as:
$$ \varvec{R = R}_{0} - (\varvec{G}_{0} + \varvec{B}_{0} )/2 $$
(4)
$$ \varvec{G = G}_{0} - (\varvec{R}_{0} + \varvec{B}_{0} )/2 $$
(5)
$$ \varvec{B = B}_{0} - (\varvec{G}_{0} + \varvec{R}_{0} )/2 $$
(6)
$$ \varvec{Y = }(\varvec{G}_{0} + \varvec{B}_{0} )/2 - (\varvec{G}_{0} - \varvec{B}_{0} )/2 - \varvec{B}_{0} $$
(7)
Intensity features: use \( \varvec{L} \) to describe the brightness of the original image:
$$ \varvec{L = }(\varvec{R}_{0} + \varvec{G}_{0} + \varvec{B}_{0} )/3 $$
(8)
Direction features: the paper uses a two-dimensional Gabor filter [19] for image processing, the two-dimensional image data can be processed directly, and operational efficiency can be improved. The calculation process is as follows:
$$ \varvec{G}_{u,v} (z) = \varvec{I}\left( z \right)*\psi_{u,v} (z) $$
(9)

\( * \) is convolution operation, \( z = (x,y) \) is grey value, \( \psi_{u,v} (z) \) Defines the Gabor kernel of orientation \( u \) and scale \( v \).

The definition of multi-scale features \( \varvec{GF}(i) \), \( i = 1,2, \ldots ,5p \), \( p = u*v \). In the paper, we use \( u = 8 \), \( v = 5 \), \( p = 40 \). Using image low-level features set \( (\varvec{R}\text{,}\varvec{G}\text{,}\varvec{B}\text{,}\varvec{Y},\varvec{L}) \) and \( \psi_{u,v} (z) \) convolution operation get \( \varvec{GF}(i) \), \( (i = 1,2, \ldots ,200) \).

According to \( \varvec{GF}(i) \) and region covariance (Sect. 2.1), each pixel can generate 200-dimensions eigenvector \( \varvec{Q}(x,y) \) and region covariance \( \varvec{C} \).

2.3 Multi-scale Tensor Region Covariance

We begin with a known dictionary consisting of \( k \) \( n*n \) positive definite matrices \( \{ \varvec{A}_{i} \}_{i = 1}^{k} \), where each \( \varvec{A}_{i} \) is referred to as a dictionary atom. Given a positive definite matrix (region covariance) \( \varvec{C} \), our goal is to represent the new matrix as a linear combination of the dictionary atoms, as follow:
$$ \varvec{C} = \sum\limits_{i = 1}^{k} {y_{i} \varvec{A}_{i} } $$
(10)

\( y_{i} (i = 1,2, \ldots ,k) \) are the vector of coefficients.

According to the literature [20, 21], it proposed tensor sparse coding to get sparse representation of \( \varvec{C} \). We can get the corresponding value \( y_{i} \) of each pixel. The eigenvector representation is \( \varvec{\varphi }(C) = \{ y_{i} \}_{i = 1}^{k} \).

In order to calculate the distance between \( R_{i} \) and \( R_{j} \), the definitions of \( \varvec{x}_{i} \) and \( \varvec{x}_{j} \) as the center of the area \( R_{i} \) and \( R_{j} \). The distance \( G(R_{i} ,R_{j} ) \) can be calculated as follow:
$$ G(R_{i} ,R_{j} ) = ||\varvec{\varphi }(\varvec{C}_{i} ) - \varvec{\varphi }(\varvec{C}_{j} )||/(1 + ||\varvec{x}_{i} - \varvec{x}_{j} ||) $$
(11)
Then by calculating the distance with the nearest \( m \) neighboring area and getting the saliency of the region, the formula is:
$$ S(R_{i} ) = (1/m)\sum\limits_{j = 1}^{m} {G(R_{i} ,R_{j} )} $$
(12)

3 Experiment

3.1 Experimental Data Set

This paper uses the CAT2000 dataset with 20 categories of scenes [22], including: (1) action, (2) emotion, (3) art, (4) black and white, (5) cartoon, (6) irregular shape, (7) indoors, (8) drawings, (9) low resolution, (10) noisy, (11) object, (12) artificial outdoors, (13) natural outdoors, (14) pattern, (15) free, (16) satellite, (17) sketch, (18) social. Where the resolution of the image is 1920 * 1080 pixels and different categories of scenes are suitable for studying the different aspects of saliency behavior. Part of the CAT2000 data set shown in Fig. 1.
Fig. 1.

Part of the CAT2000 data set

3.2 Measure Principles

Theoretically, any model can be measured by any standard, and the saliency model produces a saliency graph S, which is usually quantified by comparison with the eye movement data G. In the literature, Judd et al. have the following measures of the saliency model [23]: (1) AUC-Judd; (2) SIM; (3) EMD; (4) AUC-Borji; (5) sAUC; (6) CC; (7) NSS; (8) KL.

3.3 Visualize Measurement Data

In order to visualize the saliency of the statistical data, the visualization of the measurement criteria data is based on the Benchmark et al. method [24]. The experimental visualization data are as follows (Fig. 2):
Fig. 2.

Input image, fixation map, fixation locations

Visualize area under ROC curve (AUC) computation: A set of AUC visualizations that plot the level sets of the saliency map at different thresholds, the corresponding true positives and false negatives and the resulting ROC curve, the results are shown in Fig. 3.
Fig. 3.

Visualize area under ROC curve

Visualize normalized scan path saliency (NSS) computation: NSS plotted in parula color scale; higher values mean higher normalized saliency at fixated locations, the results are shown in Fig. 4.
Fig. 4.

Visualize normalized scan path saliency (NSS) (Color figure online)

Visualize Kullback-Leibler Divergence (KL) computation: KL heat map, with brighter red values corresponding to higher KL divergence (salMap fails to sufficiently approximate fixMap at those pixels), the results are shown in Fig. 5.
Fig. 5.

Visualize Kullback-Leibler Divergence (KL) (Color figure online)

Visualize Similarity (SIM) computation: histogram intersection, plotted in parula color scale; higher values correspond to higher intersection between saliency map and ground truth fixation map, the results are shown in Fig. 6.
Fig. 6.

Visualize similarity (SIM) (Color figure online)

Visualize Pearson’s Correlation Coefficient (CC) computation: correlation coefficient, plotted in parula color scale; higher values correspond to higher correlation between saliency map and ground truth fixation map, the results are shown in Fig. 7.
Fig. 7.

Visualize Pearson’s correlation coefficient (CC) (Color figure online)

Visualize Earth Mover’s Distance (EMD) computation: in green are regions of the saliency map from which density needs to be moved to the regions in red to transform salMap to match fixMap, The results are shown in Fig. 8.
Fig. 8.

Visualize earth mover’s distance (EMD) (Color figure online)

Visualize Information Gain (IG) computation: visualize information gain over baseline map cyan color scheme corresponds to gain relative to baseline (salMap better predicts fixMap at these pixels) red color scheme corresponds to loss relative to baseline (baseMap better predicts fixMap at these pixels), the results are shown in Fig. 9.
Fig. 9.

Visualize information gain (IG) (Color figure online)

3.4 Model Comparison

By comparing the saliency of the CAT2000 data set, the saliency maps are obtained and compared with the other algorithms. As shown in Fig. 10, an image is selected from each type of image set. Color feature as an important feature of the image, the algorithm focuses on the use of color features, the test images are color images, excluding the gray images of CAT2000 data set.

Table 1 and Fig. 10 show the results of the comparison with other up-to-date methods on the CAT2000 dataset. In this paper, we compare with the following algorithms: Goferman and others [25] proposed CAS (PAMI), which presented a detection algorithm which is based on four principles observed in the psychological literature, the benefits of the proposed approach are evaluated in two applications where the context of the dominant objects is just as essential as the objects themselves; Erdem and others [26] proposed CovSal, they proposed to use covariance matrices of simple image features (known as region covariance descriptors in the computer vision community; Hage and others [6] based on the local image block different from the surrounding image block to establish the image visual saliency (model GBVS), which uses the Markov chain for measuring the heterogeneity between local blocks; Fang and others [27] proposed to estimate image saliency (model LDS)by learning a set of discriminative subspaces that perform the best in popping out targets and suppressing distractors; Riche and others [28] proposed a novel saliency prediction model, called RARE2012, which selects information worthy of attention based on multi-scale spatial rarity; Murray and others [29] proposed an efficient model of color appearance in human vision, which contains a principled selection of parameters as well as an innate spatial pooling mechanism, can be generalized to obtain a saliency model that outperforms state-of-the-art models; Ran and others [30] proposed an approach for saliency detection that combines previously suggested patch distinctness with an object probability map(model CGI2012).
Table 1.

Correlation results by using the CAT2000 data set.

Model name

AUC-Judd

SIM

EMD

AUC-Borji

CC

NSS

KL

Judd Model

0.84

0.46

3.60

0.84

0.54

1.30

0.94

GBVS

0.80

0.51

2.99

0.79

0.50

1.23

0.80

CAS

0.77

0.50

3.09

0.76

0.42

1.07

1.04

IttiKoch

0.56

0.34

4.66

0.53

0.09

0.25

6.71

Murray model

0.70

0.43

3.79

0.70

0.30

0.77

1.14

CGI2012

0.77

0.46

4.17

0.76

0.43

1.14

1.53

RARE2012

0.82

0.54

2.72

0.81

0.57

1.44

0.76

LDS

0.83

0.58

2.09

0.79

0.62

1.54

0.79

The paper

0.87

0.46

1.78

0.79

0.36

1.89

2.62

Fig. 10.

The saliency map comparison with other up-to-date algorithms.

As shown in Fig. 10 and Table 1, the interior of the original image region is relatively uniform, and the extraction of the saliency feature is relatively difficult. CAS (PAMI) is based on four principles observed in the psychological literature, but the emphasis of saliency feature is not obvious enough, GBVS, Murray mode and CGI2012 also encounter this problem; RARE2012 algorithm selects information worthy of attention based on multi-scale spatial rarity, due to the influence of noise, the saliency extraction is ineffective; CovSal, LDS and the algorithm proposed can be very good to obtain the clear boundary and clear features, and they are good to distinguish features and disturbances, especially the proposed algorithm has been almost targeted to the specific saliency of the object, other disturbances have been excluded, the saliency effect is very obvious.

4 Conclusion

Now the most of algorithms dealt with the low-level higher-order data by vectorization, the structure of the data may be lost in the process, which leads to a poorer representation and overall performance degradation. In this paper we use regional covariance to extract image saliency features, and we introduce an approach for sparse representation of region covariance which will preserve the inherent structure of the image. Firstly it obtains the image low-level data (color, intensity, direction, etc.), and then it uses the multi-scale filtering to process all of the low-level data, on the one hand it filters out the noises of the low-level features, on the other hand it obtains image features for tensor sparse coding from the multi-band perspective, it is advantageous to preserve the inherent structure of the multi-band from the image. At last, it compares the experimental results with the commonly used feature extraction algorithms. The experimental results show that the proposed algorithm is closer to the actual boundary of the object and achieving better results.

References

  1. 1.
    Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: Proceedings of the IEEE Conference on Computer Vision Pattern Recognition (CVPR), pp. 1521–1528, June 2011Google Scholar
  2. 2.
    Borji, A., Sihite, D.N., Itti, L.: Salient object detection: a benchmark. In: Proceedings of the European Conference on Computer Vision (ECCV), pp. 414–429 (2012)Google Scholar
  3. 3.
    Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Trans. Pattern Anal. Mach. Intell. 20(11), 1254–1259 (1998)CrossRefGoogle Scholar
  4. 4.
    Koch, C., Ullman, S.: Shifts in selective visual attention: towards the underlying neural circuitry. Hum. Neurobiol. 4(4), 219–227 (1985)Google Scholar
  5. 5.
    Treisman, A.M., Gelade, G.: A feature-integration theory of attention. Cognit. Psychol. 12(1), 97–136 (1980)CrossRefGoogle Scholar
  6. 6.
    Harel, J., Koch, C., Perona, P.: Graph-based visual saliency. In: Proceedings of the Advances in Neural Information Processing Systems (NIPS), pp. 542–552 (2006)Google Scholar
  7. 7.
    Liu, T., et al.: Learning to detect a salient object. IEEE Trans. Pattern Anal. Mach. Intell. 33(2), 353–367 (2011)CrossRefGoogle Scholar
  8. 8.
    Bruce, N.D.B., Tsotsos, J.K.: Saliency based on information maximization. In: Proceedings of the Advances in Neural Information Processing Systems (NIPS), pp. 155–162 (2005)Google Scholar
  9. 9.
    Hou, X., Zhang, L.: Dynamic attention: searching for coding length increments. In: Proceedings of the Advances in Neural Information Processing Systems (NIPS), pp. 681–688 (2008)Google Scholar
  10. 10.
    Zhang, L., Tong, M.H., Marks, T.K., Shan, H., Cottrell, G.W.: SUN: a Bayesian framework for saliency using natural statistics. J. Vis. 8(7), 1–20 (2008)CrossRefGoogle Scholar
  11. 11.
    Kong, L., Duan, L., Yang, W., Dou, Y.: Salient region detection: an integration approach based on image pyramid and region property. IET Comput. Vis. 9(1), 85–97 (2015)CrossRefGoogle Scholar
  12. 12.
    Fang, Y., Lin, W., Chen, Z., Tsai, C.-M., Lin, C.-W.: A video saliency detection model in compressed domain. IEEE Trans. Circuits Syst. Video Technol. 24(1), 27–38 (2014)CrossRefGoogle Scholar
  13. 13.
    Jiang, H., Wang, J., Yuan, Z., Liu, T., Zheng, N., Li, S.: Automatic salient object segmentation based on context and shape prior. In: Proceedings of the British Machine Vision Conference (BMVC), pp. 1–12 (2011)Google Scholar
  14. 14.
    Cerf, M., Frady, E.P., Koch, C.: Faces and text attract gaze independent of the task: experimental data and computer model. J. Vis. 9(12), 1–15 (2009)CrossRefGoogle Scholar
  15. 15.
    Judd, T., Ehinger, K., Durand, F., Torralba, A.: Learning to predict where humans look. In: Proceedings of the IEEE 12th International Conference on Computer Vision (ICCV), pp. 795–825, September/October 2009Google Scholar
  16. 16.
    Chang, K.-Y., Liu, T.-L., Chen, H.-T., Lai, S.-H.: Fusing generic abjectness and visual saliency for salient object detection. In: Proceedings of the International Conference on Computer Vision (ICCV), pp. 914–921, November 2011Google Scholar
  17. 17.
    Alexe, B., Deselaers, T., Ferrari, V.: What is an object? In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 73–80, June 2010Google Scholar
  18. 18.
    Tuzel, O., Porikli, F., Meer, P.: Region covariance: a fast descriptor for detection and classification. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952, pp. 589–600. Springer, Heidelberg (2006). doi: 10.1007/11744047_45 CrossRefGoogle Scholar
  19. 19.
    Daugman, J.G.: Uncertainty relation for resolution in space, spatial frequency, and orientation optimized by two-dimensional visual cortical filters. J. Opt. Soc. Am. A 2(7), 1160–1169 (1985)CrossRefGoogle Scholar
  20. 20.
    Wang, S., Cheng, B., Ye, J., Wang, M.: Method of face image feature extract based on weighted multi-scale tensor subspace. J. Data Acquis. Process. 31(04), 791–7 98 (2016)Google Scholar
  21. 21.
    Sivalingam, R., Boley, D., Morellas, V., Papanikolopoulos, N.: Tensor sparse coding for region covariances. In: Daniilidis, K., Maragos, P., Paragios, N. (eds.) ECCV 2010. LNCS, vol. 6314, pp. 722–735. Springer, Heidelberg (2010). doi: 10.1007/978-3-642-15561-1_52 CrossRefGoogle Scholar
  22. 22.
    Borji, A., Itti, L.: CAT2000: a large scale fixation dataset for boosting saliency research. In: CVPR 2015 Workshop on “Future of Datasets” (2015). arXiv preprint arXiv:1505.03581
  23. 23.
    Bylinskii, Z., Judd, T., Oliva, A., Torralba, A., Durand, F.: What do different evaluation metrics tell us about saliency models? arXiv preprint arXiv:1604.03605 (2016)
  24. 24.
    Judd, T., Durand, F., Torralba, A.: A benchmark of computational models of saliency to predict Hum. Fixat. (2012)Google Scholar
  25. 25.
    Goferman, S., Zelnik-Manor, L., Tal, A.: Context-aware saliency detection. IEEE Trans. Pattern Anal. Mach. Intell. 34(10), 1915–1926 (2012)CrossRefGoogle Scholar
  26. 26.
    Erdem, E., Erdem, A.: Visual saliency estimation by nonlinearly integrating features using region covariances. J. Vis. 13(4), 103–104 (2013)CrossRefGoogle Scholar
  27. 27.
    Fang, S., Li, J., Tian, Y., et al.: Learning discriminative subspaces on random contrasts for image saliency analysis. IEEE Trans. Neural Netw. Learn. Syst. (2016)Google Scholar
  28. 28.
    Riche, N., Mancas, M., Duvinage, M., et al.: RARE2012: a multi-scale rarity-based saliency detection with its comparative statistical analysis. Sig. Process. Image Commun. 28(6), 642–658 (2013)CrossRefGoogle Scholar
  29. 29.
    Murray, N., Vanrell, M., Otazu, X., et al.: Saliency estimation using a non-parametric low-level vision model. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2011), Colorado Springs, 20–25 June, DBLP, pp. 433–440 (2011)Google Scholar
  30. 30.
    Ran, M., Zelnik-Manor, L., Tal, A.: Saliency for image manipulation. Vis. Comput. 29(5), 381–392 (2013)CrossRefGoogle Scholar

Copyright information

© Springer International Publishing AG 2017

Authors and Affiliations

  • Shimin Wang
    • 1
    Email author
  • Mingwen Wang
    • 1
  • Jihua Ye
    • 1
  • Anquan Jie
    • 1
  1. 1.College of Computer Information and EngineeringJiangxi Normal UniversityNanchangChina

Personalised recommendations