This application claims the benefit, under 35 U.S.C. §365 of International Application PCT/EP2007/054195, filed Apr. 27, 2007, which was published in accordance with PCT Article 21(2) on Nov. 8, 2007 in English and which claims the benefit of European patent application No. 06300418.8, filed Apr. 28, 2006 and European patent application No. 06300538.3, filed May 31, 2006.
The present invention relates to a method for estimating the salience of an image, and more particularly to a salience estimation method for object-based visual attention model.
As a neurobiological conception, attention implies the concentration of mental powers upon an object by close or careful observation. Attention area is the area in a picture where tends to catch more human attention. The system designed to automatically detect the attention area of a picture is called attention model. The detected attention area is widely utilized in many kinds of applications, such as accumulating limited resource in an attention area, directing retrieval/search, simplifying analysis, etc.
Different from attention models used in most previous machine vision systems which drive attention based on the spatial location hypothesis with macro-block (MB) being the basic unit, other models which direct visual attention are object-driven, called object-based visual attention model.
A lot of researches on MB (macro-block) spatial-based visual attention are established as proposed by L. Itti et al., “A Model of Salience-Based Visual Attention for Rapid Scene Analysis”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Volume 20, No. 11, November 1998 and by Y. F. Ma et al., “A User Attention Model for Video Summarization”, ACM Multimedia'02, pp. 533-542, December 2002. However, object-based visual attention is not so widely studied because of its inherent difficulty. Y. Sun et al. propose a framework of object-based visual attention in “Object-based Visual Attention for Computer Vision”, Artificial Intelligence, pp. 77-123, May 2003. Another object-based visual attention model is presented by F. Orabona et al., “Object-based Visual Attention: a Model for a Behaving Robot”, 3rd International Workshop on Attention and Performance in Computational Vision, June 2005. Both object-based visual attention schemes still follow the general architecture of attention model listed in
No matter in MB spatial-based or in object-based visual attention models, low level spatial/temporal features are first extracted, and then for each salient (different, outstanding from its surroundings; or say, more attractive) feature map of each unit is estimated over the whole picture, after that a master “salience map” is generated by feeding all feature maps in a purely bottom-up manner.
Compared with object-based visual attention model, the MB spatial-based visual attention model is a much easier and faster creation. However, it has several inherent disadvantages:
1) The attention area breaks natural object boundary;
2) Each micro-block may cover lots of natural objects.
So, the extracted feature of the micro-block is a mixed property of all these natural objects and thus will lower down attention area detection precision.
The key issue of the object-based visual attention model lies in two aspects: one is the object grouping before feature extraction, the other is the particular efficient salience estimation of each object over all the objects in the image. The central idea of the currently used salience estimation scheme is based on Gauss distance measure as presented by Y. Sun et al.
Denote x as the object to be salience estimated, yi (i=1, 2, . . . , n) as all the background objects, was the maximum of the width and height of the input image, and //x−yi// as the physical distance between x and yi, so the Gauss distance is defined as the formula (1),
with the scale σ set to w/ρ, where ρ is a positive integer and generally 1/ρ may be set to a percentage of w such as 2%, 4%, 5% or 20%, 25%, 50%, etc.
Denote SF(x, yi) as the absolute difference of object x and yi in feature F, then the salience estimation SF(X) as the overall salience degree of object x in feature F can be expressed as Formula (2).
By the definition of the salience estimation, it can be concluded that:
1. The larger difference between the object and its surroundings exists, the more salient the object is.
2. The closer the object and its feature differed surroundings is, the more salient the object is. That is, human vision decreases its ability to distinguish the difference according to distance. The attenuation coefficient is measured by dgauss, which is coherent with the visual physiology thesis.
This guarantees SF(x) is a useful salient estimation in feature F. Unfortunately, some important human perception properties are not considered in SF(x).
a is an original image of Skating to be estimated and
b is an original image of Coastguard to be estimated and
Both in
From
Also in
From forgoing description we can see that the conventional object-based visual attention model is not efficient enough and there are a lot of human vision properties not considered:
1. Object size—The estimation of the influence that the object size on salience degree is a complex problem. For example, (a) if all neighbouring objects yi are of the same size s and the size of object x decreases from s to 0, as a result the salience degree of x (SF(x)) will decrease gradually, (b) if all neighbouring objects yf are of the same size s and the size of object x decreases from s1 to s2 (s1>>s, and s1>s2>s), SF(X) will increase gradually. Thus we know that the relationship between object size and salience degree is not monotonous. And the problem becomes even more complex when each of the objects may have an arbitrary size.
2. Local effect—If an object is not salient among its near neighbours (local area) while the far neighbours are greatly different from the object, there are two possible results: (a) the object is not salient at all inside the whole image; (b) the local area as a whole is salient inside the image with the object being a member of the local area. No matter in which case, the salient degree of the object does not match what defined above.
3. Video texture—Suppose the object features of an image are uniformly random, human will usually ignore the details of the whole image and not any object of the image is salient, while the above defined SF(x) will be a large number for any of the objects in the image.
With all these limitations, the conventional object-based visual attention model is far from applicable. Therefore an improved object-based visual attention model is desirable.
The present invention provides a salience estimation scheme for object-based visual attention model employing a multi-level concentric circled scheme capable of lowering the computing complexity and being more applicable.
In one aspect, the invention provides a method for estimating the salience of an image. It comprises steps of segmenting the image into a plurality of objects to be estimated; extracting feature maps for each segmented object; calculating the saliences of each segmented object in a set of circles defined around a centre pixel of the object based on the extracted feature maps; and integrating the saliences of each segmented object in the all circles in order to achieve an overall salience estimation for each segmented object. According to one preferred embodiment, the step of extracting feature maps is based on the measure of image colour variation. According to another preferred embodiment, the step of calculating the salience of each segmented object comprises a sub-step of comparing colour features of the object to be estimated with that of any other object in each circle defined around the object to be estimated.
Advantageously, the object-based visual attention model based on multi-level concentric circled salience estimation scheme of the present invention presents an efficient framework to construct object-based visual attention model, which is of low computing complexity and much more human vision inosculated.
Other characteristics and advantages of the invention will be apparent through the description of a non-limiting embodiment of the invention, which will be illustrated with the help of the accompanying drawings.
a illustrates an original image of Skating to be salience estimated;
b illustrates an original image of Coastguard to be salience estimated;
a is the salience estimation result of
b is the salience estimation result of
a is an example of segmentation result of
b is another example of segmentation result of
a illustrates the estimated salience result of
b illustrates the estimated salience result of
The technical features of the present invention will be described further with reference to the embodiments. The embodiments are only preferable examples without limiting to the present invention. It will be well understood by the following detail description in conjunction with the accompanying drawings.
From the foregoing description we can see that the salience estimation process can be denoted as:
Input: An image I={pi|i=1 . . . w*h} with width w and height h;
Output: Salience map sal[1 . . . w*h] with sal[i] the salient degree of pixel pi in the image.
The method of the present invention mainly includes three steps described as below:
Step 1—Pre-processing (image segmentation)
An image I is first decomposed into a set of objects I={o1, o2, . . . , on} in this step. Image-based segmentation and grouping play a powerful role in human vision perception, a lot of researches have been developed in this area. In this invention we adopt the object segmentation scheme proposed by P. F. Felzenszwalb et al., “Image Segmentation Using Local Variation”, IEEE Computer Society on Computer Vision and Pattern Recognition, June 1998, which is based on measures of image colour variation. Here gives a simple description of this scheme.
Before processing, an undirected graph H=(V, E) is defined based on the image I, where each pixel pi of I has a corresponding vertex viεV, and an edge (vi, vj)εE connects vertices vi and vj. The precise definition of which pixels are connected by edges in E depends on the expression (1-1).
E={(vi,vj)|∥pi−pj∥≦d} (1-1)
For some given distance d., a weight function on the edges, weight(·), provides some non-negative measures of similarity (or difference) between the individual vertices vi and vj. Define weight(·) as expression (1-2),
where colour (vi) is the colour of pixel pi in the image.
S={Ci} denotes a segmentation of V and each Ci corresponds to a segmented object. Define the internal variation of C as formula (I-3),
where MST(C, E) is a minimum spanning tree of C with respect to the set of E.
Define the external variation of two objects C1 and C2 as formula (1-4).
The process of the segmentation is to make the expression (1-5) satisfied for any two of the segmented objects:
Ext(C1,C2)≦min(Int(C1)+k/|C1|,Int(C2)+k/|C2|) (1-5)
where k is a constant number set to 100 in our implementation. Denote Int(C)+k/|C| as the extended internal variation inside object C.
To achieve the segmented result, E is first sorted into n=(e1, e2, . . . , em) by non-decreasing edge weight and initially segment the image into w*h single pixel objects, then for each eq=(vi,vj) (q=1, 2, . . . , m) repeat the next process: if vi and vj belong to different objects and weight (vi, vj) is not larger than the extended internal variation(Int(C)+k/|C|) of the two objects they belong to, the two objects are merged to form a new single object.
It can be seen, this gives an efficient object segmentation scheme which will not cost too much of computing resource. In implementation, here uses an 8-connected neighbourhood for constructing E, that is d=1.
Step 2—Pre-Processing (Feature Extraction)
With yi, ri, gi, bi denoting the luminance, red, green and blue channels of pixel pi, following we extract the features of each object segmented in step 1.
Considering the definition of extended internal variation Int(C)+k/|C|, wherein k/|C| is an addition to the internal variation because underestimating the internal variation is bad for preventing components from growing. As a result, small objects are more likely to grow regardless of the internal variation inside. For example, in the Skating example of
To solve the feature extraction problem in the mentioned situation, an operator Major(f, o) is defined in the feature map F=f(vi) of an object o={v1, v2, . . . , vt}. The returned value of Major (f, o) is the representative feature of the object o which is defined to satisfy (d1, d2 and η are constant number, set to 2, 64 and 95% respectively in our implementation):
(1) If there exists a range [min, max] meeting the expression max−min+1≦d1 and the percentage of features whose values inside the range [min, max] over the whole feature map F=(f(vi), f(v2), . . . , f(vt)) are not smaller than η, Major(f, o) is then defined as the average value of those features whose values are inside the range [min, max].
(2) Otherwise, if an object size is larger than a constant d2, the object is divided into two sub-objects by the same process as in step 1 and then loop above step for each sub-object; otherwise if the object is too small, just define Major(f, o) as the average value of all the features.
With operator Major (f, o), Yi, Ri, Gi, Bi and Yei being defined as the luminance, red, green, blue and yellow channels of object of (negative values are set to 0):
Yi=Major(y,oi)
Ri=Major(r,oi)−(Major(g,oi)+Major(b,oi))/2
Gi=Major(g,oi)−(Major(r,oi)+Major(b,oi))/2
Bi=Major(b,oi)−(Major(r,oi)+Major(g,oi))/2
Yei=(Major(r,oi)+Major(g,oi))/2−|Major(r,oi)−Major(g,oi)|/2−Major(b,oi)
The intensity feature is extracted as formula (2-1).
Ii=Yi (2-1)
The “colour double-opponent” based colour features are extracted as formula (2-2) and (2-3).
RGi=Ri−Gi (2-2)
BYi=B−Yei (2-3)
Orientation will be a certain complex feature in object based visual attention model. Since all the objects are segmented according to colour variations, the object itself then will not contain any orientation information except the border of the object. Because of this special property of the segmented objects, we will not consider orientation in the implementation.
Comparing with orientation, motion will be a more possible additional feature since currently optical flow techniques become more and more mature.
But for simplicity, we only consider the three feature maps Ii, RGi and BYi in present invention.
Step 3 Salience Estimation
After the above two steps the image I is segmented into objects I=(o1, o2, . . . , on) and three feature maps Ii, RGi and BYi (i=1 . . . n) are extracted. The remaining problem is how to estimate the salience map for each feature map F (F⊂{I, RG, BY}), denoted as SalF(oi).
For any object oi of the image, denote s, as the size (the number of pixels inside the object) and ci=(Xi, Yi) as the centre pixel of the object. Xi and Yi are described as formula (3-1).
During the salience estimation process, each pixel of the oi is indistinctively considered equal to the center pixel ci, so the object is considered duplicated si copies of the center pixel as shown in
Based on this assumption, there presents a multi-level concentric circled scheme for salience estimation of oi. In the first step of this scheme, there defines a set of concentric circles circled around the center pixel ci of the object, C1 . . . Ct (Cj is an ellipse with horizontal radius rx and vertical radius ry, and is called Level j circle) are distributed from the near neighbouring areas of the center pixel ci to the far neighbouring areas. For each level j circle, estimate the salience of oi inside Cj, denoted as SalCF(oi, Cj), and the overall estimated salience of oi is then defined as formula (3-2), where kt is a constant number for linear integration.
Then, given an area Cj and an object oi in Cj with feature F extracted over Cj, how to estimate the salience of oi inside Cj considering human vision properties? Here we first give the definition of the operation SalCF:
(1) Set S as the set of objects with center pixel inside Cj.
(2) For each object ou in S, define F′u=abs(Fu−Fi). Then use the follow formula to calculate the weighted average of F′u in S.
(3) Define ρ as the percentage of pixels in S with F′u not larger than avgF′:
where bool(exp) returns 1 when exp is a true determinant else returns 0.
(4) With the definition of a detection function texture (·) as shown in
SalCF(oi,Cj)=avgF′×texture(ρ) (3-5)
Where texture(·) is an empirical function of ρ for detection of “audience area”, i.e. the area with random featured objects such as audience, which is more expected not to be recognized as attention. The detection function texture (ρ) satisfies that the lower the value of ρ is, the bigger the value of texture (ρ) will be, and thus the more chance this area is recognized as an “audience area” i.e. the video texture of the image. By using this detection function texture (·) there will be lower probability that the non-attention objects in the area are recognized as attention.
From the description above we can conclude the salience estimation scheme as below:
(a) For each object oi of the image, define a set of concentric circles Cj (j=1 . . . t).
(b) Calculate SalCF(oi, Cj) according to above definition of SalCF.
(c) Integrate the salience estimation for all Cj according to expression (3-2) to get the overall estimated salience.
a and
Whilst there has been described in the forgoing description preferred embodiments and aspects of the present invention, it will be understood by those skilled in the art that many variations in details of design or construction may be made without departing from the present invention. The present invention extends to all features disclosed both individually, and in all possible permutations and combinations.
The present object-based visual attention model based on multi-level concentric circled salience estimation scheme gives a more accuracy on understanding of the image and a far more computing efficiency, it has several advantages as below:
1. The invention presents an efficient framework to construct object-based visual attention model. It is of low computing complexity.
2. The presented framework is much more human vision inosculated. The un-considered human vision properties in conventional schemes (such as object size, local effect and video texture) are well issued.
3. The framework is extendable.
Number | Date | Country | Kind |
---|---|---|---|
06300418 | Apr 2006 | EP | regional |
06300538 | May 2006 | EP | regional |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/EP2007/054195 | 4/27/2007 | WO | 00 | 10/17/2008 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2007/125115 | 11/8/2007 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
6282317 | Luo et al. | Aug 2001 | B1 |
7940985 | Sun et al. | May 2011 | B2 |
20050084136 | Xie et al. | Apr 2005 | A1 |
Number | Date | Country |
---|---|---|
1017019 | Jul 2000 | EP |
Entry |
---|
J. Luo et al: “On measuring low-level self and relative saliency in photographic images” Pattern Recognition Letters, vol. 22, No. 2, Feb. 2001, pp. 157-169, XP004315118. |
L. Itti et al.: “A Saliency-Based Search Mechanism for Overt and Covert Shifts of Visual Attention”, Vision Research, vol. 40, No. 10-12, Jun. 2000, pp. 1489-1506, XP008060077. |
Search Report Dated Jun. 25, 2007. |
Number | Date | Country | |
---|---|---|---|
20090060267 A1 | Mar 2009 | US |