A Novel Coarse-to-Fine Hair Segmentation Method - Semantic Scholar
A Novel Coarse-to-Fine Hair Segmentation Method Dan Wang1, 2, Xiujuan Chai1, Hongming Zhang3, Hong Chang1, Wei Zeng3, Shiguang Shan1 1
Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (CAS), Beijing 100190, China 2 Graduate University of CAS, Beijing 100190, China 3 NEC Labs China, Beijing, 100084, China {dwang, xjchai, hchang, sgshan}@jdl.ac.cn; {zengwei, zhanghongming}@nec.cn
Abstract—Segmenting hair regions from human images facilitates many tasks like hair synthesis and hair style trends forecast. However, hair segmentation is quite challenging due to hair/background confusion and large hair pattern diversity. To address these problems to some extent, this paper proposes a novel coarse-to-fine hair segmentation method. In our approach, firstly, the recently proposed “Active Segmentation with Fixation” (ASF) is used to coarsely define an enclosed candidate region with high-recall (but possibly low-precision) of hair pixels and exclude considerable part of the backgrounds which are easily confused with hair. Then Graph Cuts (GC) method is applied to the candidate regions to remove additional false positives by incorporating hair-specific information. Specifically, Bayesian method is employed to select some reliable hair and background regions (seeds) among the ones over-segmented by Mean Shift. SVM classifier is then learnt online from these seeds and explored to predict hair/background likelihood probability, which is subsequently fed into GC algorithm. The novelty of the proposed approach lies in three folds: 1) an elaborate design of hair segmentation framework, which utilizes ASF to reduce the candidate hair regions and adopts GC to achieve more accurate hair region contours; 2) the region-based strategy for seed selection; 3) the exploration of the discriminative method, SVM, to predict the probability of each pixel belonging to hair and background regions. Extensive experimental results demonstrate the approach outperforms recently proposed methods.
literature dedicated to hair segmentation. In an early work [9], the authors segment hair regions in a pixel classification manner using textural and geometrical features. In [5], hair is segmented by mainly employing a hair color model, the performance of which is only evaluated subjectively. Another recent work in [10] applies matting method based on frequential and color information. Approaches in [5] and [10] both need to first find some pixels (seeds) that are surely in hair region, which, however, is not easy. Two recent methods with better segmentation performance are presented in [11] and [12]. In [11], Bayesian method is utilized for hair seed selection and segmentation, but the background information is neglected and the setting of threshold parameters is non-trivial. GC algorithm is explored in [12] for hair segmentation. Nevertheless, global optimal solution of GC is not necessarily consistent with optimal hair segmentation according to human visual perception, since there are too many unpredictable complex background regions involved in the GC optimization. In summary, previous works cannot deal with hair segmentation under the scene with complex and scattered background.
Since hair is a crucial element of human appearance, hair acquisition, styling, animation and synthesis have attracted growing interest in recent years [1-4]. Other than these graphical applications, human hair also plays a significant role in computer vision applications like human recognition [5], because people do not change hair styles frequently. In addition, hairstyle information is utilized for gender classification in the literature [6]. To accomplish all of the above tasks, extracting features from hair region is valuable and necessary. In most previous literature, however, hair regions were generally assumed segmented perfectly or manually labeled beforehand [7, 8]. Therefore, automatically segmenting hair within images is still an open problem.
In this paper, we elegantly design a novel coarse-to-fine hair segmentation method. In the proposed approach, ASF [13], which was much recently proposed to find the enclosing contour around the object fixation, is first performed to coarsely define a candidate region with high-recall but possible low-precision of hair pixels. Considerable backgrounds, which are easily confused with hair, are then removed after ASF. Subsequently, GC is applied to the candidate region to exclude additional false positives by incorporating image-specific hair/background information. Before performing GC, Bayesian based seed selection method is explored, which is similar to the one in [11]. The difference is that we perform seed selection in the form of over-segmented regions, rather than pixels. After that, SVM classifier [14] is learnt online from hair and background seed pixels and employed to predict hair likelihood probability, for succedent GC. The main difference between our approach and [12] lies in three main respects. Firstly, we propose a coarse-to-fine hair segmentation framework, which improve the segmentation accuracy significantly. Secondly, we apply a region-based strategy for seed selection, which can provide more seeds while preserving the accuracy. Thirdly, we explore SVM for hair pixel prediction, which is able to utilize discriminative information of hair/backgrounds.
Hair segmentation is a difficult task primarily due to the diversity of hair patterns and the confusion between hair and complex background. To our best knowledge, there is limited
The paper is organized as follows. A brief overview of the proposed method is given in Section II. The region-based seed selection by Bayesian rule and the online SVM learning is
Keywords-Hair segmentation; Coarse-to-fine; Graph Cuts; Active segmentation with fixation; Bayesian method
I.
INTRODUCTION
Input image Face/Head detection
Normalization
Mean Shift segmentation
and operated in pixel-wise, which can thus correct the error of Mean Shift to some extent and ensure the quantity of training samples.
Train color and location model (Offline) Generic hair color model
Region-based Bayesian seed selection Hair seed selection
Hair occurrence prior probability
Pixel-based Graph Cuts
Hair-labeled image
Background seed selection
A fixation point
(b)
(e)
(f)
(c)
(d)
ASF
SVM online learning and predicting SVM classifier
Figure 1. Flowchart illustrating the framework of the proposed method.
presented in Section III. In Section IV, we introduce the details of the coarse-to-fine hair segmentation methods including ASF and GC algorithm. Section V presents experimental results of our method and compares with other state-of-the-art methods followed by the conclusion Section VI. II.
(a)
(g)
(h)
Figure 2. An example of hair segmentation procedure: (a) An input image; (b) The over-segmentation image by Mean Shift; (c) Hair seeds (red) and background seeds(blue) selected by Bayesian method; (d) Enclosed contour by ASF; (e) SVM output probability of predicting each pixel being hair; (f) The fused probability of SVM and HOPP (Hair Occurrence Prior Probability); (g) The segmentation result; (h) Final hair segmentation result overlapped on the input image.
OVERVIEW OF THE PROPOSED METHOD
In this section, we briefly give an overview the framework of the proposed coarse-to-fine hair segmentation method, which integrates coarse segmentation by ASF and accurate segmentation by GC. As shown in Fig. 1, given an image, e.g. Fig. 2(a), we first detect faces (we detect heads instead for multi-pose hair segmentation case1) for image normalization. Then Mean Shift [15] is conducted to over-segment the image into hundreds of homogenous regions as shown in Fig. 2(b). After that, color and location information are combined into a Bayesian model for hair and background seed selection based on the oversegmented regions (as shown in Fig. 2(c)). Hair/background seeds mean the selected pixels or regions that are surely belonging to hair/background regions. Once these seed regions are obtained, on the one hand, image-specific hair/background SVM classifier can be learnt from the pixels within these seed regions; on the other hand, one of those hair seed pixels is selected as the “fixation point” fed into ASF step to produce the coarse enclosing contour (as exampled in Fig. 2(d)) enveloping hair candidate regions. Finally, GC is applied within the hair candidate region (the small background pixels at the lower right corner are removed by the ASF step, as shown in Fig. 2(d)) to achieve the final result (as shown in Fig. 2(g) and (h)), based on the combination probability (Fig. 2(f)) of the output probability from SVM classifier (referring to Fig. 2(e)) and the hair occurrence prior probability (HOPP) of location (Fig. 2(e) and (f)). It should be noted that although Bayesian method is applied on regions, the online hair/background SVM classifier is learnt Due to privacy problem, the head detection method cannot be introduced in this paper.
Figure 3. Visualization of multi-pose HOPP according to head location. Lighter gray level indicates higher probability of hair. The results in this figure are learned from 2000 head-shoulder images randomly selected from MHD3820 database, described below.
III.
HAIR SEED SELECTION AND IMAGE-SPECIFIC SVM CLASSIFIER
In this section, we will first introduce how to select the initial hair/background seeds by applying Bayesian model to the over-segmented regions produced by Mean Shift. Then we describe the online learning of image-specific SVM classifier of hair/background. A. Region-based Seed Selection by Bayesian Rule 1) Offline Learning of Hair Location and Color Prior Model Intuitively, human hair is generally located around the face, but with higher probability above face. Therefore, the prior probability of hair location relative to face can be an important cue for hair segmentation. As in [11], we call this prior by hair occurrence prior probability (HOPP). We learn it from a prepared face image dataset with hair region labeled. To calculate HOPP for each location, all of the hair-labeled images are firstly normalized to the same size according to two-eye
(for near-frontal case) or the head rectangle location (for nonfrontal case). Fig. 3 shows the visualization of multi-pose hair prior models according to normalized head rectangle location. In the offline training process, the generic hair color model (GHCM) [11], which is represented by Gaussian Mixture Model (GMM) distribution, is learnt from an existing hair database using Expectation-Maximization (EM) algorithm. 2) Region-Based Seed Selection With the learned two prior probabilities, i.e., GHCM and HOPP, the seed selection is formulated as a Bayesian model. Unlike [11] which performs pixel-wise Bayesian model, our model operates in region-wise mode based on the oversegmented regions. This region-based strategy can greatly reduce Bayesian posterior calculations, while preserving the accuracy. Formally, let f(.) denote RGB vector of a region R or a pixel x and S(.) denote the label, with 1 for hair and 0 for background. We denote HOPP of a region R or pixel x by P(S(.) = 1), and denote the conditional probability under GHCM by P(f(.)|S(.) = 1). Thus, according to Bayesian rules, posterior probability of a region R being a hair region can be computed by: P S R 1 | f R
P f R | S R 1 P S R 1 P f R
,
model like GMM used in [11] and [12], the advantage of SVM resides in its ability to explore more discriminative information of hair and backgrounds. The output probability of SVM classifier for each pixel is fed into GC for accurate hair segmentation, which will be described in the next section. IV.
COARSE-TO-FINE HAIR SEGMENTATION
Hair segmentation is challenging due to two main reasons: hair/background confusion and hair pattern diversity. Therefore one of the key problems is how to remove more complicated backgrounds while preserving the true positive pixels. ASF is capable of fulfilling both of these requirements, by providing a coarse hair candidate region. Then accurate segmentation is performed within this scope using image-specific, onlinelearned information. The advantage of this coarse-to-fine strategy is: ASF can remove considerable complex backgrounds which are easily confused with neighboring hair; GC can be performed only within the defined candidate region and thus can avoid being significantly interfered by unknown regions far from hair.
(1)
where
f ( R)
1 . f ( x) , R xR
P S R 1
1 . P S x 1 , R xR
(2) (3)
and P S x 1 is the HOPP of the pixel location x. After computing hair posterior probability for each region according to (1), the regions with highest probabilities are selected as hair seeds and the regions with the lowest probabilities are appointed as background seeds. The region-based Bayesian seed selection performance will be compared with pixel-based one in [11] later in experimental section. B. Hair/Background SVM Online Learning With above-selected hair and background seed regions, it is a natural idea to learn a discriminant model to capture the image-specific discriminative information of hair and backgrounds. Unlike the statistical models such as GMM used in literature [11] and [12], which only contributes to hair and background description independently, SVM is explored in our method to better differentiate hair from backgrounds,. A Support Vector Machine (SVM) is a learning algorithm for pattern classification and regression [14]. The basic training principle behind SVM is finding the optimal linear hyper-plane such that the expected classification error for unseen test samples is minimized. In our implementation, libsvm package is used [16]. In our SVM classifier, RGB vector is adopted as the input feature and the training samples are hair and background seed pixels, the selection method of which has been elaborated in Subsection A. Compared with the statistical
Figure 4. ASF results examples
A. Active Segmentation with Fixation (ASF) ASF has been recently proposed by Ajay Mishra et al [13]. The basic idea of the method is that a fixation always lies inside a region in the image and the region boundary encloses an object or just a part of an object. Thus the method to segment an object in an image is to find the enclosing contour around the “fixation point”. Our motivation to employ ASF is that most hair regions appear to be relatively uniform patterns and connected with face and neck regions. Hair regions can be supposed to be segmented as an object or a part of other objects such as headshoulder, providing a “fixation point” of hair regions is fed into the ASF step. If the premise is established, the whole hair regions maybe segmented, or at least a great deal of complex backgrounds can be rejected. Some ASF results are shown in Fig. 4 and detailed experimental results in Section V demonstrates that ASF can produce hair candidate regions with high-recall, while removing considerable confusing complex backgrounds. B. Hair Segmentation by Pixel-Based GC With the selected seed regions, image-specific hair/background SVM classifier has been learnt online and the hair/background probability for each pixel in the image can be predicted from SVM. In this subsection, we will present how to
construct GC energy function. In the graph, each pixel is used as a graph node for segmentation. Extracting hair from an image with various backgrounds can be formulated as a binary labeling problem. Let A A1 ,..., Ak ,... A be a binary vector whose components Ak
specifies assignment to pixel Rk in the pixel set . The pixel Rk can be either hair or background. Vector A defines a hair/background labeling or segmentation and each Ak is 1 for a hair pixel and 0 for background. An energy function is formulated as: E A D A B A ,
(4)
where is used to control relative importance of the two terms. The first term D is called data term or regional term, which reflects to what extent the color and location properties of pixel Rk fit into the observed hair or background seed regions. Here, the Bayesian posterior probability D(A) is utilized to describe how well an input pixel fit into hair or background model:
D A P f Rk | Ak P Ak .
(5)
k 1
The Bayesian model is similar to that of the seed selection in Section 0. The difference is that P f Rk | Ak is the SVM output probability of each pixel being hair or background, rather than the GHCM likelihood probability. Besides, P Ak is the value of the HOPP in each pixel location. The second term in (4) is called the smoothness term or boundary term, which can be interpreted as a penalty for a discontinuity between neighboring pixels. 8-neighbor configuration is used here. We choose a function related to the color feature distance:
B A
Rp , Rq Νe ighbors
B R p , Rq Ap , Aq ,
(6)
where， 1, if Ap Aq , 0, otherwise
Ap , Aq
f R f R p q B R p , Rq exp 2 2
2
1 , Dist ( R , R ) p q
(7)
(8)
where f(R) is the RGB vector of the pixel R and is obtained by calculating the average color distance of neighboring pixels, which reflects the level of variation between neighboring regions. Minimization of the energy function can be solved via the min-cut/max-flow algorithm as described in [18] and the solution corresponds to hair segmentation results. It should be noted that pixel-based, rather than region-based, GC is adopted since the over-segmentation boundaries are not always consistent with the enclosed contour generated from ASF.
V.
EXPERIMENTAL RESULTS
In this section, experiments are performed in three aspects: firstly, we evaluate the performance of Bayesian method for seed selection, where our region-based strategy is compared with the pixel-based method in [11]; secondly, we validate the effects of ASF; finally, the proposed coarse-to-fine hair segmentation method is compared with the approaches in [11], [12] and the latest work [19] which can reliably segment the foreground object contour from complex background by maximal similarity based region merging (MSRM) strategy. A. Database In our experiments, we collect two databases from Internet for performance evaluation (partly from [10]). One database, named as “Near-frontal Head-shoulder Database 1000” (NHD1000), contains 1000 near-frontal head-shoulder images, with 50% female and 50% male images respectively. This database covers a large range of variations in hair styles, hair colors and backgrounds. The ground truth of hair regions are labeled manually and the images are normalized to size 80 100 with 16 pixels for two-eye distance. The other database is more challenging, named as “Multi-pose Headshoulder Database 3820” (MHD3820). The database includes 3820 head-shoulder images. There are eight different poses in MHD3820 images, in which 2900 images are near-frontal, and the rest 920 are non-frontal. The images in MHD3820 are normalized to 80 100 according to head rectangle positions. It should be noted that this database includes 172 Caltech human images used in [10]. This database contains much more cluttered backgrounds than NHD1000 and includes more challenging illustration variations, complex indoor and outdoor environments, and diverse hair colors. All experiments below need the prior models of GHCM and HOPP. Thus we collect another 300 images from Internet to learn the models offline. These images are not overlapping with NHD1000 and MHD3820. B. Experiments on Seed Selection Experiments are conducted on NHD1000, to compare our region-based Bayesian seed selection method with the pixelbased one. For both methods, we adopt the GHCM and HOPP mentioned in the previous subsection. The number of GMM components is empirically set to 13. In Bayesian framework, the weights for GHCM and HOPP are set to 0.7 and 0.3 respectively by experience. Besides, for our region-based strategy, the criterion to set the parameters of Mean Shift is that each over-segmented region contains either hair or non-hair pixels only. Here we set the spatial and color parameters to 5 and 2 respectively and the minimum region contains at least 50 pixels. The criterion for an image owning correct seeds is its precision of the selected hair seed pixels being up to 95%. We use the percentage of images with correct seeds to evaluate the performances. Experiments tell, on the premise of ensuring 90% correct rate, 370 hair pixel seeds are achieved by our region-based Bayesian method, while the pixel-based one can only produce 320 hair seeds. While preserving the same accuracy, more seeds can definitely be used to learn more
robust SVM classifier, leading to more accurate segmentation results. TABLE I.
EVALUATION OF THE EFFECT OF ASF
Measurements
NHD1000
MHD3820
precision
46%
38%
recall
95%
94%
C. Experiments on ASF In this subsection, we evaluate the effectiveness and robustness of ASF method for hair segmentation. As mentioned in Section IV, an input “fixation point” is needed for ASF. Here the central pixel of the hair seed region is adopted. Since ASF is expected to provide a coarse contour enclosing most hair regions, we evaluate the recall rate of hair pixels within the contour the segmentation on both NHD1000 and MHD3820 databases. TABLE I. depicts the ASF with a proper input “fixation point” can produce enclosed contour of hair pixels with high-recall rate. Besides, ASF is insensitive to the “fixation point”. As shown in Fig. 5, although different fixation points are input into ASF, the extracted contours are relatively stable.
Figure 5. ASF results with different fixation point.
D. Experiments on Hair Segmentation The proposed method is compared with the Bayesian method in [11], the MSRM strategy in [19] and GC approach in [12] on both of near-frontal and non-frontal head-shoulder images. All of these methods need the seed selection step. For fair comparison, the proposed region-based Bayesian method is utilized to select seeds for all of these methods in our experiments. For MSRM, the source code can be obtained from website [20], which needs human interactive control. Here, instead of using manually labeled seed regions for MSRM, we use the automatically selected ones to make it an automatic method.
The proposed algorithm is evaluated by assessing the consistency of segmentation results and the manually labeled ground truth in terms of the F-measure [21], defined as 2PR/(P+R), where P stands for the precision which calculates the hair pixels of automatic segmentation overlapping with the ground truth, and R stands for recall which measures the hair pixels of the ground-truth segmentation overlapping with automatic segmentation. As F-measure is defined at all points on the precision-recall curve, the maximum F-measure value for each method is reported in this paper. TABLE II. compares the performance of the proposed method and the approaches in [11], [19] and [12]. In these
experiments, parameters have been adjusted for the compared methods to ensure their respective best performances. For example, for the Bayesian method [11], the probability threshold, the number of GMM components is tuned, while the for both GC method and the proposed coarse-to-fine method is tuned. It can be observed that the proposed coarse-to-fine strategy achieves superior performance to other methods on both databases. Some visualized results of these methods are given in column (b-e) in Fig. 6, in which the bottom row is a failure example and others are successful examples with the proposed method. Note that, column (f) just shows the ASF result, not the final segmentation. The two main reasons contributing to successful coarse-to-fine segmentation are: first, the ASF usually reduces the hair candidate regions and removes many complicated backgrounds; second, GC can produce satisfactory segmentation in the enclosed contour, within which there are not too many confusing backgrounds. The problem for Bayesian method is that it is impossible to tune parameters for all images and it does not consider the consistency of neighboring pixels. MSRM is a general region merging segmentation method. The disadvantage of MSRM is that it must be assured that each of the unconnected parts of the foreground is provided with some input seed pixels, or the foreground cannot be segmented completely (Row 3). The difficulty for GC method is tuning adaptive for different images, since it can never be predicted how the backgrounds effect the optimization. Compared with the original GC, the proposed method conquers such problem to some extent, as it adopts ASF to remove considerable backgrounds and thus avoid certain background influences on GC optimization. Nevertheless, there are also some failure cases for the proposed method, such as the bottom row in Fig. 6. The coarse enclosed contour, generated from the ASF failed to cover all the hair regions, due to the confusion of hair and dark clothes. Other failure cases include images with complex illumination and hair scattering in several discontinuous regions. It should be noted that, since multi-pose database MHD3820 includes non-frontal images, in which eyes cannot be located, we use head rectangle for normalization procedure. Experimental results show that the proposed method is robust to the variations of head-shoulder poses and complicated indoor and outdoor backgrounds. VI.
CONCLUSIONS
This paper presents a novel coarse-to-fine hair segmentation method. The ASF method is first employed to coarsely define an enclosed candidate region with high-recall/low-precision of hair pixels; while GC is subsequently explored inside the candidate region to reject additional false positives by combining online-learned hair/background SVM classifiers and location prior properties. The main advantages of our proposed method are: 1) ASF can coarsely define a hair candidate region with high-recall rate, within which performing GC can be prevented from being interfered by unpredictable complicated backgrounds and thus produce more accurate segmentation; 2) region-based seed selection can provide more seeds while preserving accuracy; 3) prediction probability of SVM considers discriminative information of hair/background.
TABLE II.
PERFORMANCE EVALUATION OF DIFFERENT METHODS
Algorithm
NHD1000 Precision/ F-score Recall
MHD3820 Precision/ F-score Recall
Bayesian method [11]
0.656
0.589/0.806 0.640
0.736/0.623
MSRM [19]
0.740
0.772/0.783 0.697
0.808/0.697
GC based approach[12] 0.813
0.868/0.808 0.669
0.823/0.637
Proposed method
0.887/0.848 0.773
0.788/0.817
0.850
fine strategy can be further applied to more general forebackground segmentation problems in the future. ACKNOWLEDGMENT This paper is partially supported by Natural Science Foundation of China under contracts No.60833013, No.61001193, and No.60872077; National Basic Research Program of China (973 Program) under contract 2009CB320902; and Hi-Tech Research and Development Program of China under contract No.2009AA01Z317. REFERENCES [1]
[2]
[3]
[4] [5]
[6]
[7] [8]
[9]
[10] [11]
[12]
[13]
[14] [15] (a)
(b)
(c)
(d)
(e)
(f)
Figure 6. Results of applying our method (e) compared to other latest segmentation algorithms (b-d). (a) Original image; (b) Bayesian method [11] ; (c) MSRM [19]; (d) GC [12]; (e) The proposed method; (f) ASF results.
The experimental results demonstrate the superiority of the proposed method for hair segmentation, which coarsely limit the search scope firstly and then segment hair within the candidate region with GC method accurately. The coarse-to-
[16] [17] [18]
[19]
[20] [21]
S. Hadap, N. Magnenat-Thalmann, “Modeling dynamic hair as continuum,” Computer Graphics Forum, vol. 20, pp. 329–338, September 2001. S. Paris, H. Briceño, F. Sillion, “Capture of hair geometry from multiple images,” ACM SIGGRAPH. New York, vol. 23, pp. 712–719, August 2004. F. Bertails, B. Audoly, B. Querleux, F. Leroy, J. Leveque, M. Cani, “Predicting natural hair shapes by solving the statics of flexible rods,” Eurographics. Trinity College, Dublin, Irlande, August 2005. K. Ward, N. Galoppo, M. Lin, “Interactive virtual hair salon,” MIT Press, vol. 16, pp. 237-251, June 2007. Y.Yacoob and L. S. Davis, “Detection and analysis of Hair,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 28, pp. 11641169, July 2006. K. Ueki, H. Komatsu, S. Imaizumi, K. Kaneko, N. Sekine, J. Katto, T. Kobayashi, “A method of gender classification by integrating facial, hairstyle, and clothing images,” ICPR, vol. 4, pp. 446-449, August 2004. L. Wang, Y. Yu, K. Zhou, B. Guo, “Example-based hair geometry synthesis,” ACM SIGGRAPH, New York, vol. 28, August 2009. S. Paris, W. Chang, O. I. Kozhushnyan, W. Jarosz, W. Matusik, M. Zwicker, F. Durand, “Hair photobooth: geometric and photometric acquisition of real hairstyles,” ACM SIGGRAPH, New York, vol. 27, August 2008. Z. Liu, J. Guo, L. Bruton, “A knowledge-based system for hair region segmentation,” International Symposium on Signal Processing and Its Applications. Australia, vol. 2, pp. 575-576, August 1996. C. Rousset and P.Y. Coulon, “Frequential and color analysis for hair mask segmentation,” ICIP. San Diego, pp. 2276-2279, October 2008. D. Wang, S. Shan, W. Zeng, H. Zhang, X. Chen, “A novel two-tier Bayesian based method for hair segmentation,” ICIP. Cairo, Egypt, pp. 2377-2380, November 2009. K. Lee, D. Anguelov, B. Sumengen, B. Gokturk, “Markov random field models for hair and face segmentation,” IEEE International Conference on Automatic Face and Gesture Recognition, pp. 1-6, September 2008. A. Mishra, Y. Aloimonos, C. L. Fah, “Active segmentation with fixation,” International Conference on Computer Vision. Kyoto, pp. 468-475, October 2009. C. Cortes, V. Vapnik, “Support vector networks” Machine Learning. Boston, vol. 20, pp. 273-297, 1995. D. Comaniciu and P. Meer, “Mean Shift: a robust approach toward feature space analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 28, pp. 603–619, May 2002. http://www.csie.ntu.edu.tw/~cjlin/libsvm/ http://www.umiacs.umd.edu/~mishraka/activeSeg.html Y. Boykov, O. Veksler, and R. Zabih, “Fast approximate energy minimization via graph cuts,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 23, pp. 1222-1239, November 2001. J. Ning, L. Zhang, D. Zhang, C. Wu, “Interactive image segmentation by maximal similarity based region merging”, Pattern Recognition, vol. 43, pp. 445-456, February 2010. http://www4.comp.polyu.edu.hk/~cslzhang/papers.htm C. J. Van Rijsbergen, Information Retrieval, 2nd ed., 1979.
Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (CAS), Beijing 100190, China 2 Graduate University of CAS, Beijing 100190, China 3 NEC Labs China, Beijing, 100084, China {dwang, xjchai, hchang, sgshan}@jdl.ac.cn; {zengwei, zhanghongming}@nec.cn
Abstract—Segmenting hair regions from human images facilitates many tasks like hair synthesis and hair style trends forecast. However, hair segmentation is quite challenging due to hair/background confusion and large hair pattern diversity. To address these problems to some extent, this paper proposes a novel coarse-to-fine hair segmentation method. In our approach, firstly, the recently proposed “Active Segmentation with Fixation” (ASF) is used to coarsely define an enclosed candidate region with high-recall (but possibly low-precision) of hair pixels and exclude considerable part of the backgrounds which are easily confused with hair. Then Graph Cuts (GC) method is applied to the candidate regions to remove additional false positives by incorporating hair-specific information. Specifically, Bayesian method is employed to select some reliable hair and background regions (seeds) among the ones over-segmented by Mean Shift. SVM classifier is then learnt online from these seeds and explored to predict hair/background likelihood probability, which is subsequently fed into GC algorithm. The novelty of the proposed approach lies in three folds: 1) an elaborate design of hair segmentation framework, which utilizes ASF to reduce the candidate hair regions and adopts GC to achieve more accurate hair region contours; 2) the region-based strategy for seed selection; 3) the exploration of the discriminative method, SVM, to predict the probability of each pixel belonging to hair and background regions. Extensive experimental results demonstrate the approach outperforms recently proposed methods.
literature dedicated to hair segmentation. In an early work [9], the authors segment hair regions in a pixel classification manner using textural and geometrical features. In [5], hair is segmented by mainly employing a hair color model, the performance of which is only evaluated subjectively. Another recent work in [10] applies matting method based on frequential and color information. Approaches in [5] and [10] both need to first find some pixels (seeds) that are surely in hair region, which, however, is not easy. Two recent methods with better segmentation performance are presented in [11] and [12]. In [11], Bayesian method is utilized for hair seed selection and segmentation, but the background information is neglected and the setting of threshold parameters is non-trivial. GC algorithm is explored in [12] for hair segmentation. Nevertheless, global optimal solution of GC is not necessarily consistent with optimal hair segmentation according to human visual perception, since there are too many unpredictable complex background regions involved in the GC optimization. In summary, previous works cannot deal with hair segmentation under the scene with complex and scattered background.
Since hair is a crucial element of human appearance, hair acquisition, styling, animation and synthesis have attracted growing interest in recent years [1-4]. Other than these graphical applications, human hair also plays a significant role in computer vision applications like human recognition [5], because people do not change hair styles frequently. In addition, hairstyle information is utilized for gender classification in the literature [6]. To accomplish all of the above tasks, extracting features from hair region is valuable and necessary. In most previous literature, however, hair regions were generally assumed segmented perfectly or manually labeled beforehand [7, 8]. Therefore, automatically segmenting hair within images is still an open problem.
In this paper, we elegantly design a novel coarse-to-fine hair segmentation method. In the proposed approach, ASF [13], which was much recently proposed to find the enclosing contour around the object fixation, is first performed to coarsely define a candidate region with high-recall but possible low-precision of hair pixels. Considerable backgrounds, which are easily confused with hair, are then removed after ASF. Subsequently, GC is applied to the candidate region to exclude additional false positives by incorporating image-specific hair/background information. Before performing GC, Bayesian based seed selection method is explored, which is similar to the one in [11]. The difference is that we perform seed selection in the form of over-segmented regions, rather than pixels. After that, SVM classifier [14] is learnt online from hair and background seed pixels and employed to predict hair likelihood probability, for succedent GC. The main difference between our approach and [12] lies in three main respects. Firstly, we propose a coarse-to-fine hair segmentation framework, which improve the segmentation accuracy significantly. Secondly, we apply a region-based strategy for seed selection, which can provide more seeds while preserving the accuracy. Thirdly, we explore SVM for hair pixel prediction, which is able to utilize discriminative information of hair/backgrounds.
Hair segmentation is a difficult task primarily due to the diversity of hair patterns and the confusion between hair and complex background. To our best knowledge, there is limited
The paper is organized as follows. A brief overview of the proposed method is given in Section II. The region-based seed selection by Bayesian rule and the online SVM learning is
Keywords-Hair segmentation; Coarse-to-fine; Graph Cuts; Active segmentation with fixation; Bayesian method
I.
INTRODUCTION
Input image Face/Head detection
Normalization
Mean Shift segmentation
and operated in pixel-wise, which can thus correct the error of Mean Shift to some extent and ensure the quantity of training samples.
Train color and location model (Offline) Generic hair color model
Region-based Bayesian seed selection Hair seed selection
Hair occurrence prior probability
Pixel-based Graph Cuts
Hair-labeled image
Background seed selection
A fixation point
(b)
(e)
(f)
(c)
(d)
ASF
SVM online learning and predicting SVM classifier
Figure 1. Flowchart illustrating the framework of the proposed method.
presented in Section III. In Section IV, we introduce the details of the coarse-to-fine hair segmentation methods including ASF and GC algorithm. Section V presents experimental results of our method and compares with other state-of-the-art methods followed by the conclusion Section VI. II.
(a)
(g)
(h)
Figure 2. An example of hair segmentation procedure: (a) An input image; (b) The over-segmentation image by Mean Shift; (c) Hair seeds (red) and background seeds(blue) selected by Bayesian method; (d) Enclosed contour by ASF; (e) SVM output probability of predicting each pixel being hair; (f) The fused probability of SVM and HOPP (Hair Occurrence Prior Probability); (g) The segmentation result; (h) Final hair segmentation result overlapped on the input image.
OVERVIEW OF THE PROPOSED METHOD
In this section, we briefly give an overview the framework of the proposed coarse-to-fine hair segmentation method, which integrates coarse segmentation by ASF and accurate segmentation by GC. As shown in Fig. 1, given an image, e.g. Fig. 2(a), we first detect faces (we detect heads instead for multi-pose hair segmentation case1) for image normalization. Then Mean Shift [15] is conducted to over-segment the image into hundreds of homogenous regions as shown in Fig. 2(b). After that, color and location information are combined into a Bayesian model for hair and background seed selection based on the oversegmented regions (as shown in Fig. 2(c)). Hair/background seeds mean the selected pixels or regions that are surely belonging to hair/background regions. Once these seed regions are obtained, on the one hand, image-specific hair/background SVM classifier can be learnt from the pixels within these seed regions; on the other hand, one of those hair seed pixels is selected as the “fixation point” fed into ASF step to produce the coarse enclosing contour (as exampled in Fig. 2(d)) enveloping hair candidate regions. Finally, GC is applied within the hair candidate region (the small background pixels at the lower right corner are removed by the ASF step, as shown in Fig. 2(d)) to achieve the final result (as shown in Fig. 2(g) and (h)), based on the combination probability (Fig. 2(f)) of the output probability from SVM classifier (referring to Fig. 2(e)) and the hair occurrence prior probability (HOPP) of location (Fig. 2(e) and (f)). It should be noted that although Bayesian method is applied on regions, the online hair/background SVM classifier is learnt Due to privacy problem, the head detection method cannot be introduced in this paper.
Figure 3. Visualization of multi-pose HOPP according to head location. Lighter gray level indicates higher probability of hair. The results in this figure are learned from 2000 head-shoulder images randomly selected from MHD3820 database, described below.
III.
HAIR SEED SELECTION AND IMAGE-SPECIFIC SVM CLASSIFIER
In this section, we will first introduce how to select the initial hair/background seeds by applying Bayesian model to the over-segmented regions produced by Mean Shift. Then we describe the online learning of image-specific SVM classifier of hair/background. A. Region-based Seed Selection by Bayesian Rule 1) Offline Learning of Hair Location and Color Prior Model Intuitively, human hair is generally located around the face, but with higher probability above face. Therefore, the prior probability of hair location relative to face can be an important cue for hair segmentation. As in [11], we call this prior by hair occurrence prior probability (HOPP). We learn it from a prepared face image dataset with hair region labeled. To calculate HOPP for each location, all of the hair-labeled images are firstly normalized to the same size according to two-eye
(for near-frontal case) or the head rectangle location (for nonfrontal case). Fig. 3 shows the visualization of multi-pose hair prior models according to normalized head rectangle location. In the offline training process, the generic hair color model (GHCM) [11], which is represented by Gaussian Mixture Model (GMM) distribution, is learnt from an existing hair database using Expectation-Maximization (EM) algorithm. 2) Region-Based Seed Selection With the learned two prior probabilities, i.e., GHCM and HOPP, the seed selection is formulated as a Bayesian model. Unlike [11] which performs pixel-wise Bayesian model, our model operates in region-wise mode based on the oversegmented regions. This region-based strategy can greatly reduce Bayesian posterior calculations, while preserving the accuracy. Formally, let f(.) denote RGB vector of a region R or a pixel x and S(.) denote the label, with 1 for hair and 0 for background. We denote HOPP of a region R or pixel x by P(S(.) = 1), and denote the conditional probability under GHCM by P(f(.)|S(.) = 1). Thus, according to Bayesian rules, posterior probability of a region R being a hair region can be computed by: P S R 1 | f R
P f R | S R 1 P S R 1 P f R
,
model like GMM used in [11] and [12], the advantage of SVM resides in its ability to explore more discriminative information of hair and backgrounds. The output probability of SVM classifier for each pixel is fed into GC for accurate hair segmentation, which will be described in the next section. IV.
COARSE-TO-FINE HAIR SEGMENTATION
Hair segmentation is challenging due to two main reasons: hair/background confusion and hair pattern diversity. Therefore one of the key problems is how to remove more complicated backgrounds while preserving the true positive pixels. ASF is capable of fulfilling both of these requirements, by providing a coarse hair candidate region. Then accurate segmentation is performed within this scope using image-specific, onlinelearned information. The advantage of this coarse-to-fine strategy is: ASF can remove considerable complex backgrounds which are easily confused with neighboring hair; GC can be performed only within the defined candidate region and thus can avoid being significantly interfered by unknown regions far from hair.
(1)
where
f ( R)
1 . f ( x) , R xR
P S R 1
1 . P S x 1 , R xR
(2) (3)
and P S x 1 is the HOPP of the pixel location x. After computing hair posterior probability for each region according to (1), the regions with highest probabilities are selected as hair seeds and the regions with the lowest probabilities are appointed as background seeds. The region-based Bayesian seed selection performance will be compared with pixel-based one in [11] later in experimental section. B. Hair/Background SVM Online Learning With above-selected hair and background seed regions, it is a natural idea to learn a discriminant model to capture the image-specific discriminative information of hair and backgrounds. Unlike the statistical models such as GMM used in literature [11] and [12], which only contributes to hair and background description independently, SVM is explored in our method to better differentiate hair from backgrounds,. A Support Vector Machine (SVM) is a learning algorithm for pattern classification and regression [14]. The basic training principle behind SVM is finding the optimal linear hyper-plane such that the expected classification error for unseen test samples is minimized. In our implementation, libsvm package is used [16]. In our SVM classifier, RGB vector is adopted as the input feature and the training samples are hair and background seed pixels, the selection method of which has been elaborated in Subsection A. Compared with the statistical
Figure 4. ASF results examples
A. Active Segmentation with Fixation (ASF) ASF has been recently proposed by Ajay Mishra et al [13]. The basic idea of the method is that a fixation always lies inside a region in the image and the region boundary encloses an object or just a part of an object. Thus the method to segment an object in an image is to find the enclosing contour around the “fixation point”. Our motivation to employ ASF is that most hair regions appear to be relatively uniform patterns and connected with face and neck regions. Hair regions can be supposed to be segmented as an object or a part of other objects such as headshoulder, providing a “fixation point” of hair regions is fed into the ASF step. If the premise is established, the whole hair regions maybe segmented, or at least a great deal of complex backgrounds can be rejected. Some ASF results are shown in Fig. 4 and detailed experimental results in Section V demonstrates that ASF can produce hair candidate regions with high-recall, while removing considerable confusing complex backgrounds. B. Hair Segmentation by Pixel-Based GC With the selected seed regions, image-specific hair/background SVM classifier has been learnt online and the hair/background probability for each pixel in the image can be predicted from SVM. In this subsection, we will present how to
construct GC energy function. In the graph, each pixel is used as a graph node for segmentation. Extracting hair from an image with various backgrounds can be formulated as a binary labeling problem. Let A A1 ,..., Ak ,... A be a binary vector whose components Ak
specifies assignment to pixel Rk in the pixel set . The pixel Rk can be either hair or background. Vector A defines a hair/background labeling or segmentation and each Ak is 1 for a hair pixel and 0 for background. An energy function is formulated as: E A D A B A ,
(4)
where is used to control relative importance of the two terms. The first term D is called data term or regional term, which reflects to what extent the color and location properties of pixel Rk fit into the observed hair or background seed regions. Here, the Bayesian posterior probability D(A) is utilized to describe how well an input pixel fit into hair or background model:
D A P f Rk | Ak P Ak .
(5)
k 1
The Bayesian model is similar to that of the seed selection in Section 0. The difference is that P f Rk | Ak is the SVM output probability of each pixel being hair or background, rather than the GHCM likelihood probability. Besides, P Ak is the value of the HOPP in each pixel location. The second term in (4) is called the smoothness term or boundary term, which can be interpreted as a penalty for a discontinuity between neighboring pixels. 8-neighbor configuration is used here. We choose a function related to the color feature distance:
B A
Rp , Rq Νe ighbors
B R p , Rq Ap , Aq ,
(6)
where， 1, if Ap Aq , 0, otherwise
Ap , Aq
f R f R p q B R p , Rq exp 2 2
2
1 , Dist ( R , R ) p q
(7)
(8)
where f(R) is the RGB vector of the pixel R and is obtained by calculating the average color distance of neighboring pixels, which reflects the level of variation between neighboring regions. Minimization of the energy function can be solved via the min-cut/max-flow algorithm as described in [18] and the solution corresponds to hair segmentation results. It should be noted that pixel-based, rather than region-based, GC is adopted since the over-segmentation boundaries are not always consistent with the enclosed contour generated from ASF.
V.
EXPERIMENTAL RESULTS
In this section, experiments are performed in three aspects: firstly, we evaluate the performance of Bayesian method for seed selection, where our region-based strategy is compared with the pixel-based method in [11]; secondly, we validate the effects of ASF; finally, the proposed coarse-to-fine hair segmentation method is compared with the approaches in [11], [12] and the latest work [19] which can reliably segment the foreground object contour from complex background by maximal similarity based region merging (MSRM) strategy. A. Database In our experiments, we collect two databases from Internet for performance evaluation (partly from [10]). One database, named as “Near-frontal Head-shoulder Database 1000” (NHD1000), contains 1000 near-frontal head-shoulder images, with 50% female and 50% male images respectively. This database covers a large range of variations in hair styles, hair colors and backgrounds. The ground truth of hair regions are labeled manually and the images are normalized to size 80 100 with 16 pixels for two-eye distance. The other database is more challenging, named as “Multi-pose Headshoulder Database 3820” (MHD3820). The database includes 3820 head-shoulder images. There are eight different poses in MHD3820 images, in which 2900 images are near-frontal, and the rest 920 are non-frontal. The images in MHD3820 are normalized to 80 100 according to head rectangle positions. It should be noted that this database includes 172 Caltech human images used in [10]. This database contains much more cluttered backgrounds than NHD1000 and includes more challenging illustration variations, complex indoor and outdoor environments, and diverse hair colors. All experiments below need the prior models of GHCM and HOPP. Thus we collect another 300 images from Internet to learn the models offline. These images are not overlapping with NHD1000 and MHD3820. B. Experiments on Seed Selection Experiments are conducted on NHD1000, to compare our region-based Bayesian seed selection method with the pixelbased one. For both methods, we adopt the GHCM and HOPP mentioned in the previous subsection. The number of GMM components is empirically set to 13. In Bayesian framework, the weights for GHCM and HOPP are set to 0.7 and 0.3 respectively by experience. Besides, for our region-based strategy, the criterion to set the parameters of Mean Shift is that each over-segmented region contains either hair or non-hair pixels only. Here we set the spatial and color parameters to 5 and 2 respectively and the minimum region contains at least 50 pixels. The criterion for an image owning correct seeds is its precision of the selected hair seed pixels being up to 95%. We use the percentage of images with correct seeds to evaluate the performances. Experiments tell, on the premise of ensuring 90% correct rate, 370 hair pixel seeds are achieved by our region-based Bayesian method, while the pixel-based one can only produce 320 hair seeds. While preserving the same accuracy, more seeds can definitely be used to learn more
robust SVM classifier, leading to more accurate segmentation results. TABLE I.
EVALUATION OF THE EFFECT OF ASF
Measurements
NHD1000
MHD3820
precision
46%
38%
recall
95%
94%
C. Experiments on ASF In this subsection, we evaluate the effectiveness and robustness of ASF method for hair segmentation. As mentioned in Section IV, an input “fixation point” is needed for ASF. Here the central pixel of the hair seed region is adopted. Since ASF is expected to provide a coarse contour enclosing most hair regions, we evaluate the recall rate of hair pixels within the contour the segmentation on both NHD1000 and MHD3820 databases. TABLE I. depicts the ASF with a proper input “fixation point” can produce enclosed contour of hair pixels with high-recall rate. Besides, ASF is insensitive to the “fixation point”. As shown in Fig. 5, although different fixation points are input into ASF, the extracted contours are relatively stable.
Figure 5. ASF results with different fixation point.
D. Experiments on Hair Segmentation The proposed method is compared with the Bayesian method in [11], the MSRM strategy in [19] and GC approach in [12] on both of near-frontal and non-frontal head-shoulder images. All of these methods need the seed selection step. For fair comparison, the proposed region-based Bayesian method is utilized to select seeds for all of these methods in our experiments. For MSRM, the source code can be obtained from website [20], which needs human interactive control. Here, instead of using manually labeled seed regions for MSRM, we use the automatically selected ones to make it an automatic method.
The proposed algorithm is evaluated by assessing the consistency of segmentation results and the manually labeled ground truth in terms of the F-measure [21], defined as 2PR/(P+R), where P stands for the precision which calculates the hair pixels of automatic segmentation overlapping with the ground truth, and R stands for recall which measures the hair pixels of the ground-truth segmentation overlapping with automatic segmentation. As F-measure is defined at all points on the precision-recall curve, the maximum F-measure value for each method is reported in this paper. TABLE II. compares the performance of the proposed method and the approaches in [11], [19] and [12]. In these
experiments, parameters have been adjusted for the compared methods to ensure their respective best performances. For example, for the Bayesian method [11], the probability threshold, the number of GMM components is tuned, while the for both GC method and the proposed coarse-to-fine method is tuned. It can be observed that the proposed coarse-to-fine strategy achieves superior performance to other methods on both databases. Some visualized results of these methods are given in column (b-e) in Fig. 6, in which the bottom row is a failure example and others are successful examples with the proposed method. Note that, column (f) just shows the ASF result, not the final segmentation. The two main reasons contributing to successful coarse-to-fine segmentation are: first, the ASF usually reduces the hair candidate regions and removes many complicated backgrounds; second, GC can produce satisfactory segmentation in the enclosed contour, within which there are not too many confusing backgrounds. The problem for Bayesian method is that it is impossible to tune parameters for all images and it does not consider the consistency of neighboring pixels. MSRM is a general region merging segmentation method. The disadvantage of MSRM is that it must be assured that each of the unconnected parts of the foreground is provided with some input seed pixels, or the foreground cannot be segmented completely (Row 3). The difficulty for GC method is tuning adaptive for different images, since it can never be predicted how the backgrounds effect the optimization. Compared with the original GC, the proposed method conquers such problem to some extent, as it adopts ASF to remove considerable backgrounds and thus avoid certain background influences on GC optimization. Nevertheless, there are also some failure cases for the proposed method, such as the bottom row in Fig. 6. The coarse enclosed contour, generated from the ASF failed to cover all the hair regions, due to the confusion of hair and dark clothes. Other failure cases include images with complex illumination and hair scattering in several discontinuous regions. It should be noted that, since multi-pose database MHD3820 includes non-frontal images, in which eyes cannot be located, we use head rectangle for normalization procedure. Experimental results show that the proposed method is robust to the variations of head-shoulder poses and complicated indoor and outdoor backgrounds. VI.
CONCLUSIONS
This paper presents a novel coarse-to-fine hair segmentation method. The ASF method is first employed to coarsely define an enclosed candidate region with high-recall/low-precision of hair pixels; while GC is subsequently explored inside the candidate region to reject additional false positives by combining online-learned hair/background SVM classifiers and location prior properties. The main advantages of our proposed method are: 1) ASF can coarsely define a hair candidate region with high-recall rate, within which performing GC can be prevented from being interfered by unpredictable complicated backgrounds and thus produce more accurate segmentation; 2) region-based seed selection can provide more seeds while preserving accuracy; 3) prediction probability of SVM considers discriminative information of hair/background.
TABLE II.
PERFORMANCE EVALUATION OF DIFFERENT METHODS
Algorithm
NHD1000 Precision/ F-score Recall
MHD3820 Precision/ F-score Recall
Bayesian method [11]
0.656
0.589/0.806 0.640
0.736/0.623
MSRM [19]
0.740
0.772/0.783 0.697
0.808/0.697
GC based approach[12] 0.813
0.868/0.808 0.669
0.823/0.637
Proposed method
0.887/0.848 0.773
0.788/0.817
0.850
fine strategy can be further applied to more general forebackground segmentation problems in the future. ACKNOWLEDGMENT This paper is partially supported by Natural Science Foundation of China under contracts No.60833013, No.61001193, and No.60872077; National Basic Research Program of China (973 Program) under contract 2009CB320902; and Hi-Tech Research and Development Program of China under contract No.2009AA01Z317. REFERENCES [1]
[2]
[3]
[4] [5]
[6]
[7] [8]
[9]
[10] [11]
[12]
[13]
[14] [15] (a)
(b)
(c)
(d)
(e)
(f)
Figure 6. Results of applying our method (e) compared to other latest segmentation algorithms (b-d). (a) Original image; (b) Bayesian method [11] ; (c) MSRM [19]; (d) GC [12]; (e) The proposed method; (f) ASF results.
The experimental results demonstrate the superiority of the proposed method for hair segmentation, which coarsely limit the search scope firstly and then segment hair within the candidate region with GC method accurately. The coarse-to-
[16] [17] [18]
[19]
[20] [21]
S. Hadap, N. Magnenat-Thalmann, “Modeling dynamic hair as continuum,” Computer Graphics Forum, vol. 20, pp. 329–338, September 2001. S. Paris, H. Briceño, F. Sillion, “Capture of hair geometry from multiple images,” ACM SIGGRAPH. New York, vol. 23, pp. 712–719, August 2004. F. Bertails, B. Audoly, B. Querleux, F. Leroy, J. Leveque, M. Cani, “Predicting natural hair shapes by solving the statics of flexible rods,” Eurographics. Trinity College, Dublin, Irlande, August 2005. K. Ward, N. Galoppo, M. Lin, “Interactive virtual hair salon,” MIT Press, vol. 16, pp. 237-251, June 2007. Y.Yacoob and L. S. Davis, “Detection and analysis of Hair,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 28, pp. 11641169, July 2006. K. Ueki, H. Komatsu, S. Imaizumi, K. Kaneko, N. Sekine, J. Katto, T. Kobayashi, “A method of gender classification by integrating facial, hairstyle, and clothing images,” ICPR, vol. 4, pp. 446-449, August 2004. L. Wang, Y. Yu, K. Zhou, B. Guo, “Example-based hair geometry synthesis,” ACM SIGGRAPH, New York, vol. 28, August 2009. S. Paris, W. Chang, O. I. Kozhushnyan, W. Jarosz, W. Matusik, M. Zwicker, F. Durand, “Hair photobooth: geometric and photometric acquisition of real hairstyles,” ACM SIGGRAPH, New York, vol. 27, August 2008. Z. Liu, J. Guo, L. Bruton, “A knowledge-based system for hair region segmentation,” International Symposium on Signal Processing and Its Applications. Australia, vol. 2, pp. 575-576, August 1996. C. Rousset and P.Y. Coulon, “Frequential and color analysis for hair mask segmentation,” ICIP. San Diego, pp. 2276-2279, October 2008. D. Wang, S. Shan, W. Zeng, H. Zhang, X. Chen, “A novel two-tier Bayesian based method for hair segmentation,” ICIP. Cairo, Egypt, pp. 2377-2380, November 2009. K. Lee, D. Anguelov, B. Sumengen, B. Gokturk, “Markov random field models for hair and face segmentation,” IEEE International Conference on Automatic Face and Gesture Recognition, pp. 1-6, September 2008. A. Mishra, Y. Aloimonos, C. L. Fah, “Active segmentation with fixation,” International Conference on Computer Vision. Kyoto, pp. 468-475, October 2009. C. Cortes, V. Vapnik, “Support vector networks” Machine Learning. Boston, vol. 20, pp. 273-297, 1995. D. Comaniciu and P. Meer, “Mean Shift: a robust approach toward feature space analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 28, pp. 603–619, May 2002. http://www.csie.ntu.edu.tw/~cjlin/libsvm/ http://www.umiacs.umd.edu/~mishraka/activeSeg.html Y. Boykov, O. Veksler, and R. Zabih, “Fast approximate energy minimization via graph cuts,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 23, pp. 1222-1239, November 2001. J. Ning, L. Zhang, D. Zhang, C. Wu, “Interactive image segmentation by maximal similarity based region merging”, Pattern Recognition, vol. 43, pp. 445-456, February 2010. http://www4.comp.polyu.edu.hk/~cslzhang/papers.htm C. J. Van Rijsbergen, Information Retrieval, 2nd ed., 1979.
