外文翻譯---特征空間穩(wěn)健性分析彩色圖像分割 英文

1、Robust Analysis of Feature Spaces: Color Image Segmentation Dorin Comaniciu Peter Meer Department of Electrical and Computer Engineering Rutgers University, Piscataway, NJ 08855, USA comanici @caip.rutgers.edu Abstr

2、act A general technique for the recovery o fsignificant im- age features is presented. The technique is based on the mean shift algorithm, a simple nonparametric procedure for estimating density gradients. Drawbacks

3、of the current methods (including robust clustering) are avoided. Feature space of any nature can be processed, and as an example, color image segmentation is discussed. The segmentation is completely autonomous, on

4、ly its class is chosen by the usel: Thus, the same program can produce a high quality edge image, or provide, by extracting all the significant colors, a preprocessorfor content-based query systems. A 512 x 512 color

5、 image is analyzed in less than 1 0seconds on a stan- dard workstation. Gray level images are handled as color images having only the lightness coordinate. 1. Introduction Feature space analysis is a widely used tool f

6、or solv- ing low-level image understanding tasks. Given an image, feature vectors are extracted from local neighborhoods and mapped into the space spanned by their components. Signif- icant features in the image the

7、n correspond to high density regions in this space. Feature space analysis is the procedure of recovering the centers of the high density regions, i.e., the representations of the significant image features. Histogr

8、am based techniques, Hough transform are examples of the ap- proach. To avoid the artifacts of quantization a feature space should have a continuous coordinate system. The content of a continuous feature space can

9、be modeled as a sam- ple from a multivariate, multimodal probability distribution. The highest density regions correspond to clusters centered on the modes of the underlying probability distribution. Tra- ditional c

10、lustering techniques [5], can be used but they are reliable only if the number of clusters is small and known a priori. There is no theoretical evidence that individual clus- ters obey multivariate normal distributi

11、ons and an extracted normal cluster necessarily corresponds to a significant im- age feature. On the contrary, a strong artifact cluster may appear when several features are mapped into partially over- lapping region

12、s. meer @ caip.rutgers.edu Nonparametric density estimation [4, Chap. 61 avoids the use of the normality assumption, however it requires ad- ditional input information (type of the kernel, number of neighbors). Thi

13、s information must be provided by the user, and for multimodal distributions it is difficult to guess the optimal setting. Nevertheless, a reliable general technique for feature space analysis can be developed using

14、a simple nonparametric density estimation algorithm. In this paper we propose such a technique whose robust behavior is supe- rior to methods employing robust estimators from statistics. 2. Requirements for Robustn

15、ess Estimation of a cluster center is called in statistics the multivariate location problem. To be robust, an estimator must tolerate a percentage of outliers, i.e., data points not obeying the underlying distributio

16、n of the cluster. Numer- ous robust techniques were proposed [9, Sec. 7.11, and in computer vision the most widely used is the minimum vol- ume ellipsoid ( W E ) estimator proposed by Rousseeuw [9]. Based on MVE,

17、a robust clustering technique with appli- cations in computer vision was proposed in [6]. The number of significant clusters was not needed a priori. The robust clustering method was successfully employed for the

18、analy- sis of a large variety of feature spaces, but was found to be- come less reliable once the number of modes exceeded ten. This is mainly due to constraining the shape of the removed clusters to be elliptical. F

19、urthermore, the estimated covari- ance matrices (the shape parameters of the ellipsoids) are not reliable since are based on only p + 1 points. To be able to correctly recover a large number of signif- icant f

20、eatures, the problem of feature space analysis must be solved in context. In image understanding tasks the data to be analyzed originates in the image domain. That is, the feature vectors satisfy additional, spatial

21、constraints. While these constraints are indeed used in the current techniques, their role is mostly limited to compensating for feature al- location errors made during the independent analysis of the feature space.

22、To be robust the feature space analysis must fully exploit the image domain information. As a consequence of the increased role of image domain information the burden on the feature space analysis can be reduced.

23、First all the significant features are extracted, and 1063-6919/97 $10.00 0 1997 IEEE 750 0.1741 is shown with a star at the top of Figure 1. Other, more adaptive strategies for setting the search window size can als

24、o be defined. The mean shift algorithm is the tool needed for feature space analysis. The outline of a general procedure is given below. Feature Space Analysis 1. Map the image domain into the feature space. 2. Defi

25、ne an adequate number of search windows at ran- 3. Find the high density region centers by applying the 4. Validate the extracted centers with image domain con- 5. Allocate, using image domain information, all the fea- T

26、he procedure is very general and applicable to any feature space. In the next section we describe a color image seg- mentation technique developed based on this outline. dom locations in the space. mean shift algorith

27、m to each window. straints to provide the feature palette. ture vectors to the feature palette. 4. Color Image Segmentation Image segmentation, partioning the image into homoge- neous regions, is a challenging task.

28、 The richness of vi- sual information makes bottom-up, solely image driven ap- proaches always prone to errors. To be reliable, the cur- rent systems must be large and incorporate numerous ad-hoc procedures, e.g.

29、[l]. The paradigms of gray level image segmentation (pixel-based, area-based, edge-based) are also used for color images. In addition, the physics-based meth- ods take into account information about the image forma-

30、 tion processes as well [7]. The proposed segmentation tech- nique does not consider the physical processes, it uses only the given image, i.e., a set of RGB vectors. Nevertheless, can be easily extended to incorpor

31、ate supplementary infor- mation about the input. Since perfect segmentation cannot be achieved without a top-down, knowledge driven component, a bottom-up seg- mentation technique should e only provide the input i

32、nto the next stage where the task is accomplished using a priori knowledge about its goal; and 0 eliminate, as much as possible, the dependence on user set parameter values. Segmentation resolution is the most ge

33、neral parameter char- acterizing a segmentation technique. While this parameter has a continuous scale, three important classes can be dis- tinguished. Undersegmentation corresponds to the lowest resolution. Homo

34、geneity is defined with a large tolerance margin and only the most significant colors are retained for the feature palette. The region boundaries in a correctly undersegmented image are the dominant edges in the ima

35、ge. Oversegmentation corresponds to intermediate resolution. The feature palette is rich enough that the image is bro- ken into many small regions from which any sought in- formation can be assembled under knowle

36、dge control. Oversegmentation is the recommended class when the goal of the task is object recognition. Quantization corresponds to the highest resolution. The feature palette contains all the important colors in the

37、 image. This segmentation class became important with the spread of image databases, e.g., [3, 81. The full palette, possibly together with the underlying spatial structure, is essential for content-based queries.

38、The proposed color segmentation technique operates in any of the these three classes. The user only chooses the desired class, the specific operating conditions are derived automat- ically by the program. Images

39、 are usually stored and displayed in the RGB space. However, to ensure the isotropy of the feature space, a uniform color space with the perceived color differences measured by Euclidean distances should be used. We h

40、ave chosen the L*u*v* space [lo, Sec. 3.3.91, whose coordi- nates are related to the RGB values by nonlinear transfor- mations. The daylight standard D 6 5was used as reference illuminant. The chromatic information

41、 is carried by U* and U*, while the lightness coordinate L* can be regarded as the relative brightness. Psychophysical experiments show that L*u*v* space may not be perfectly isotropic [lo, p. 3111, however, it wa

42、s found satisfactory for image understanding applications. The image capture/display operations also in- troduce deviations which are most often neglected. The steps of color image segmentation are presented be- low

43、. The acronyms ID and FS stand for image domain and feature space respectively. All feature space computations are performed in the L*u*v* space. 1. [FS] Dejinition o fthe segmentation parameters. The user only indi

44、cates the desired class of segmentation. The class definition is translated into three parameters 0 the radius of the search window, r ; e the smallest number of elements required for a signifi- 0 the smallest number

45、of contiguous pixels required for The size of the search window determines the resolution of the segmentation, smaller values corresponding to higher resolutions. Within the same segmentation class an image containi

46、ng large homogeneous regions should be analyzed at higher resolution than an image with many textured ar- eas. The simplest measure of the “visual activity” can be derived from the global covariance matrix. The squar

47、e root of its trace, 0, is related to the power of the signal (image). The radius r is taken proportional to 0. The rules defining the three segmentation class parameters are given in Table 1. cant color, “in; a s