REGION-BASED SKIN COLOR DETECTION

Rudra P. K. Poudel

, Hammadi Nait-Charif

, Jian J. Zhang

and David Liu

Media School, Bournemouth University, Weymouth House, Talbot Campus, Poole, Dorset, BH12 5BB, U.K.

Siemens Corporate Research, 755 College Road East, Princeton, NJ 08540, U.S.A.

Keywords:

Skin-color Detection, Region-based, Superpixels, Bayes Classiﬁer, Conditional Random Field.

Abstract:

Skin color provides a powerful cue for complex computer vision applications. Although skin color detection

has been an active research area for decades, the mainstream technology is based on the individual pixels.

This paper presents a new region-based technique for skin color detection which outperforms the current

state-of-the-art pixel-based skin color detection method on the popular Compaq dataset (Jones and Rehg,

2002). Color and spatial distance based clustering technique is used to extract the regions from the images,

also known as superpixels. In the ﬁrst step, our technique uses the state-of-the-art non-parametric pixel-based

skin color classiﬁer (Jones and Rehg, 2002) which we call the basic skin color classiﬁer. The pixel-based skin

color evidence is then aggregated to classify the superpixels. Finally, the Conditional Random Field (CRF)

is applied to further improve the results. As CRF operates over superpixels, the computational overhead is

minimal. Our technique achieves 91.17% true positive rate with 13.12% false negative rate on the Compaq

dataset tested over approximately 14,000 web images.

1 INTRODUCTION

Skin color provides a powerful cue in complex com-

puter vision applications such as hand tracking, face

tracking, and pornography detection. Skin color de-

tection is computationally efﬁcient yet invariant of

rotation, scaling and occlusion. These are the ma-

jor reasons for its popularity. The main challenges of

skin color detection are illumination, ethnicity back-

ground, make-up, hairstyle, eyeglasses, background

color, shadows and motion (Kakumanu et al., 2007).

Many of the skin color detection problems could be

overcome by using infrared (Socolinsky et al., 2003)

and spectral imaging (Pan et al., 2003). However,

such systems are expensive as well as cumbersome

to implement. Moreover, there are many situations

where such systems can not be used such as image

retrieval on the internet.

Most of the skin color detection methods are pixel-

based, which treat each skin or non-skin pixel individ-

ually without considering its neighbours. However,

it is natural to treat skin or non-skin as regions in-

stead of individual pixels. Surprisingly, there are only

few region-based skin detection techniques (Yang

and Ahuja, 1998), (Kruppa et al., 2002), (Jedynak

et al., 2003) and (Sebe et al., 2004). Kruppa (Kruppa

et al., 2002), Yangand Ahuja (Yang and Ahuja, 1998)

searched for elliptical skin color shape to ﬁnd the

face. Sebe (Sebe et al., 2004) used ﬁxed 3x3 pixel

patches to train a Bayesian network and Jedyank (Je-

dynak et al., 2003) smoothed the results using hidden

Markov model. This paper proposes a new technique

purely based on the concept of regions, irrespective of

the underlying geometrical shape. As such, this tech-

nique can be easily integrated into any skin detection

based system.

Our technique uses a segmentation technique

called superpixel (Moore et al., 2008) and (Ren and

Malik, 2003) to group similar color pixels together.

Then each superpixel is classiﬁed as skin or non-skin

by aggregating pixel-based evidence obtained using a

histogram based Bayesian classiﬁer similar to (Jones

and Rehg, 2002). The result is further improved with

CRF, which operate over superpixels instead of pix-

els. Even though the segmentation cost is an overhead

over the pixel-based approach, it greatly reduces the

processing cost further down the line, such as smooth-

ing with CRF. Also, aggregation of pixels into regions

helps to reduce local redundancy and the probability

of merging unrelated pixels (Soatto, 2009). Since

superpixels preserve the boundary of the objects, it

helps to achieve very accurate object segmentations

(Fulkerson et al., 2009).

The presented method not only outperforms the

current state-of-the-art pixel-based skin color detec-

tion methods but also extracts larger skin regions

301

P. K. Poudel R., Nait-Charif H., J. Zhang J. and Liu D..

REGION-BASED SKIN COLOR DETECTION.

DOI: 10.5220/0003801203010306

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2012), pages 301-306

ISBN: 978-989-8565-03-7

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

while still keeping the false-positive rate lower, pro-

viding semantically more meaningful results. This

could in turn beneﬁt higher-level vision tasks, such

as face or hand detection. Related work is discussed

in section 2; section 3 presents the proposed region-

based skin color detection technique; experiments and

results are discussed in section 4. Finally, we summa-

rize our work in section 5.

2 RELATED WORK

Skin color detection has two important parts: one is

color space selection and another is color modelling.

RGB: (Jones and Rehg, 2002), HSV: (Zhu et al.,

2004), CIE-Lab: (Kawato and Ohya, 2002), YCbCr:

(Wong et al., 2003), and normalized RGB: (Brown

et al., 2001) are popular color spaces, with RGB and

HSV being the most frequently used. CIE-Lab uni-

formly represents the color based on how two colors

differ to the human observer. HSV shows better re-

sults under varying illumination (Kakumanu et al.,

2007). However, most camera output RGB and il-

lumination variation can be eliminated by increasing

sample size (Jones and Rehg, 2002). Due to this rea-

son the RGB color space is chosen in our experiments.

Skin color modelling falls in three categories: ex-

plicitly deﬁned skin region (Peer et al., 2003), non-

parametric and parametric methods. Histogram based

Bayes classiﬁer is a popular non-parametric mod-

elling approach. Jones and Rehg (Jones and Rehg,

2002) used RGB color space and histograms based

Bayes classiﬁer and obtained 90% true positive rate

with 14.5% false positive rate on unconstrained web

images, a dataset made up of approximately 14,000

images. On parametric skin modelling technique,

mixture of Gaussian has shown the best result (Ter-

rillon et al., 2000). However, Jones and Rehg (Jones

and Rehg, 2002) showed that given enough sam-

ples, the histogram based Bayes classiﬁer technique

is slightly better than mixture of Gaussian. Neural

Network (Phung et al., 2002), self organizing map

(Brown et al., 2001), Bayesian network (Sebe et al.,

2004) and a few other methods have been used for

skin color modelling.

This paper presents a region-based skin color de-

tection method with no prior knowledge on the geo-

metric shape of the regions. The works of Yang and

Ahuja (Yang and Ahuja, 1998), Kruppa (Kruppa

et al., 2002), Jedyank (Jedynak et al., 2003) and Sebe

(Sebe et al., 2004) are the close to ours. However,

Yang and Ahuja (Yang and Ahuja, 1998) used multi-

scale segmentations to ﬁnd elliptical regions for face

detection. Hence, their model is biased toward ellipti-

(a) Original image (b) After segmentation (c) Segmentation

visualization

Figure 1: An example of superpixel segmentation.

cal objects. Kruppa (Kruppa et al., 2002) also used a

similar concept to ﬁnd the elliptical region using color

and shape information for face detection. Sebe (Sebe

et al., 2004) used 3x3 ﬁxed size pixel patches. Our

presented technique uses patches with varying sizes,

which is purely based on image evidence, i.e. skin

color in this case. Also, Jedyank (Jedynak et al.,

2003) used hidden Markov model at pixel level, while

we use conditional random ﬁelds and operate on su-

perpixel, as described in section 3.4.

3 REGION-BASED APPROACH

We argue that skin is better presented as regions rather

than individual pixels. The proposed region-based ap-

proach has four major components: basic skin classi-

ﬁer (section 3.1), extraction of regions called super-

pixels (section 3.2), superpixels classiﬁcation (section

3.3), and a smoothing procedure with CRF (section

3.4). Each step is discussed in detail below.

3.1 Basic Skin Color Classiﬁer

Any good skin color classiﬁcation method can be used

as a basic skin color classiﬁer. This paper uses the

histogram based Bayesian classiﬁer similar to that of

Jones and Rehg (Jones and Rehg, 2002), a state-of-

the-art skin color detection technique.

Learning skin and non-skin histograms: densities

of skin and non-skin color histograms are learned

from the Compaq dataset (Jones and Rehg, 2002).

The Compaq skin color dataset has approximately

4,700 skin images and 9,000 non-skin images col-

lected from free web crawling. It has images from

all ethnic groups with uncontrolled illumination and

background conditions. The number of manually la-

belled pixels is nearly 1 billion. Skin and non-skin

histograms are obtained in RGB color space with 32

bins for each color channel, exactly same to the set-

tings in Jones and Rehg (Jones and Rehg, 2002).

Equal number of skin images are randomly divided

for training and testing. Similarly, equal number of

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

302

non-skin images are randomly divided for training

and testing.

Bayesian skin classiﬁer: Naive Bayes is used to build

the skin and non-skin classiﬁer. The probability of a

color being skin s given a color c, P(s|c), is given by

P(s|c) =

P(c|s)P(s)

P(c)

(1)

where, P(c|s) is the likelihood of a givencolor c being

skin, P(s) is skin color prior and P(c) is color prior.

Similarly, the probability of a color being non-skin ns

given a color c is given by

P(ns|c) =

P(c|ns)P(ns)

P(c)

(2)

where, P(c|ns) is the likelihood of a given color c be-

ing non-skin and P(ns) prior for non-skin color. Fur-

ther P(c) could be calculated as following

P(c) = P(c|s)P(s) + P(c|ns)P(ns) (3)

P(c|s) and P(c|ns) are directly calculated from skin

and non-skin histograms. Prior probabilities: P(s)

and P(ns) can also estimated from the total number

of skin and non-skin samples in the training dataset.

However, for skin and non-skin classiﬁcation, we can

simply compare P(s|c) to P(ns|c). Using equations

(1) and (2), the ratio of P(s|c) to P(ns|c) can be sim-

pliﬁed to

P(s|c)

P(ns|c)

P(c|s)P(s)

P(c|ns)P(ns)

(4)

Equation (4) can be thresholded to produce a skin

and non-skin classiﬁcation rule. Further, P(s) and

P(ns) are also constant so this can be simpliﬁed as

follows

P(c|s)

P(c|ns)

> Θ (5)

where Θ is a constant threshold value.

In the experiments, equation (5) is used to ﬁnd the

skin and non-skin probability for pixels. The values

of P(c|s) and P(c|ns) are directly looked-up from nor-

malized skin and non-skin histograms respectively.

3.2 Superpixels

A region or collection of pixels is called a superpixel.

A ﬁve dimensional vector is used to extract the super-

pixels: three RGB color channels and two positional

coordinatesof the pixel, using the quick shift (Vedaldi

and Soatto, 2008) image segmentation algorithm. Su-

perpixels generated from this approach vary in size

and shape, hence the number of superpixels in each

image is highly dependent upon the complexity of the

image. An image with low color variation will have

Figure 2: Comparison between pixel-based (Jones and

Rehg, 2002) (middle) and region-based with CRF(right)

skin color classiﬁcation techniques.

a smaller number of superpixels than an image with

high color variation, as there is no penalty for bound-

ary violation. Generally, the concept of boundary is

not used when extracting the superpixels, however

different objects have different texture or color which

will implicitly act as boundaries. Figure 1 shows the

example of superpixels of an image. In our work we

have used the superpixel extraction library (Vedaldi

and Fulkerson, 2008) for superpixel segmentation.

3.3 Superpixel Classiﬁcation

First, the pixel based skin color classiﬁer deﬁned on

section 3.1 is used to classify the pixels of the images.

Then the probability of being skin for a given super-

pixel sp with N number of color pixels c is deﬁned as

follows

P(s|sp) =

∑

P(s|c

) (6)

Similarly the probability of being non-skin for a

given superpixel sp with N number of color pixels c

is deﬁned as follows

P(ns|sp) =

∑

P(ns|c

) (7)

REGION-BASED SKIN COLOR DETECTION

303

(a) Original image (b) Pixel-based (c) Region-based without CRF (d) Region-based with CRF

Figure 3: This example shows the advantages of the region-based approach even without CRF (see sub ﬁgures b and c). Sub

ﬁgures c and d show the failure case when CRF is applied.

3.4 Smoothing with CRF

Skin regions have varying size and shape, depending

upon the camera angle, distance from the camera and

human body factors. Hence, to obtain smooth skin re-

gions but still preserve the skin and non-skin bound-

aries, it is necessary to introduce some constraints.

CRF provides a natural way of combining pairwise

constraints. Color difference and length of bound-

ary between adjacent superpixels are used as pairwise

constraints. Optimum skin and non-skin labelling L

of all superpixels S of an image is deﬁned as follows

−log(P(L|S;ω)) = −

∑

∈S

Ψ(l

)+ω

∑

)∈E

Φ(c

, c

, s

)

(8)

where ω is the weight of pairwise constraint, E is

the set of edges of superpixel, and i and j are index

nodes in superpixel level graph of an image.

Color potential (Ψ(l

)): the color potential Ψ cap-

tures the skin and non-skin probability of superpixel

. We have used skin and non-skin probability for

superpixel directly from superpixel classiﬁcation de-

ﬁned in section 3.3 for color potential Ψ as follows

Ψ(l

) = log(P(l

)) (9)

Edge and boundary potential (Φ(c

, c

, s

))):

pairwise edge and boundary potential Φ is deﬁned

similar to those of (Fulkerson et al., 2009)

Φ(c

, c

, s

) =



L(s

, s

)

1+ ||s

− s



, [c

6= c

] (10)

where L(s

, s

) is the shared boundary length, and

||s

−s

|| is the Euclidean norm of the color difference

between s

and s

superpixels.

Only one pairwise potential is used to make the

system as simple as possible to show that treating skin

color with regions is more effective than with pixels.

This implementation has only one weighting factor ω,

which is optimized using cross validation. We use

the multi-label graph optimization library of (Boykov

et al., 2001), (Boykov and Kolmogorov, 2004) and

(Kolmogorov and Zabih, 2004) for the inference of

skin and non-skin regions. CRF graph is built on the

superpixel level hence CRF optimization is fast.

Table 1: The results of pixel-based and our region-based

technique.

Method TP FP

Jones and Rehg (2002) 90% 14.2%

Our (superpixel only) 91.44% 13.73%

Our (superpixel and CRF) 91.17% 13.12%

4 EXPERIMENTS AND RESULTS

Equal number of training and testing sets are ran-

domly chosen from the Compaq dataset (Jones and

Rehg, 2002) and same training and testing sets are

used for all experiments. The Compaq dataset has ap-

proximately 4,700 skin and 9,000 non-skin images,

freely collected from the web. Basic pixel-based skin

color classiﬁer mentioned in section 3.1 achievessim-

ilar results to those in Jones and Rehg (Jones and

Rehg, 2002). We have used RGB bin size = 32

for each channels, and threshold constant Θ = 1. It

roughly detects 90% skin color with 14.2% false pos-

itive rate.

Superpixel extraction using quick shift is con-

trolled by three parameters: (i) λ controls the trade

off between spatial and color consistency, (ii) σ con-

trols the deviation of density estimator, and (iii) τ

maximum distance in the quick shift tree. We have

used σ = 2, τ = 6, and λ = 0.9 for our experiment.

Which are chosen using grid search as there is no

explicit mechanism to preserve the skin boundaries,

with above selected parameters we have noticed that

97.43% skin pixels are correctly grouped into super-

pixels with 0.35% false positive rate. Average size of

the superpixels increases with the larger valueof τ and

σ and vice versa. Lower values of λ give importance

to spatial factor while higher values give importance

to the color value. Average size of superpixels are

larger when λ is

∼

0.5. Skin color detection depends

upon the values of the color channels, hence higher

importance is given to the color consistency in super-

pixel extraction. Also, experiments show that the skin

boundary is not well preserved with higher spatial im-

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

304

(a) Original image (b) Pixel-based (c) Region-based CRF with color in-

formation only

(d) Region-based CRF with color

and border information

Figure 4: Example shows the failure of region-based approach when only a color difference constraint is used with CRF.

portance. The average size of superpixel is 65 in our

experiments. However, the size of superpixels is not

ﬁxed and fully depends on the complexity of the im-

ages.

Table 1 shows the results comparison between

the presented region-based technique and the cur-

rent state-of-the-art pixel-based skin color detection

(Jones and Rehg, 2002) on unconstrained illumina-

tion and background. The region based technique

without CRF has 91.44% true positive rate with

13.73% false positive rate and with CRF has 91.17%

of true positive rate with 13.12% false positive rate.

Simply grouping the pixel-based evidence onto su-

perpixels increased the true positive rate by 1.44%

and decreased the false positive rate by 0.48%. This

shows treating skin as a region yields better results

than using pixels only. Both results from the region-

based techniques are better than the pixel-based tech-

nique.

The results on ﬁgure 2 show the effectiveness

of the region-based technique with CRF over pixel-

based method. Region-based technique ﬁrst groups

the skin and non-skin evidence from each pixels into

superpixels level using basic skin color classiﬁer,

which helps to remove noise. This is the main rea-

son why only grouping the pixel-based evidence into

superpixels increases the true positive rate by 1.44%

and reduces the false positive rate by 0.5% (see table

1). Also, CRF helps further extract larger smooth skin

regions by exploiting neighbouring color information

and boundary sharing between superpixels.

However, there are also some cases where region-

based technique performs worse than pixel-based

technique when we apply the CRF. Figure 3 are such

examples. Skin-like looking pixels and high bound-

ary sharing between skin and non-skin regions are the

main reason of the failure. However, we also exper-

imented using the color difference constraint only on

CRF instead of both color difference and boundary

sharing constraints and found that it performs better

when skin regions are very small and narrow. But

overall CRF with both neighbour color difference and

length of boundary sharing constraints performed bet-

ter. Figure 4 shows an example where CRF with

both neighbours color difference and length of bound-

ary sharing performs better than only with neighbours

color difference.

Skin regions do not have the same color values,

even the closest skin color pixels within superpixels

have different color values. Also, other skin-like ob-

jects exist. Hence, results can be further improved

using texture information. This is left for our future

work.

5 CONCLUSIONS

This paper presents a region-based skin color detec-

tion technique, which outperforms the current state-

of-the-art pixel-based technique. Color and spatial

distance based clustering technique is used to extract

the regions from the images, also known as superpix-

els. In the ﬁrst step, our technique uses the state-of-

the-art non-parametric pixel-based skin color classi-

ﬁer (Jones and Rehg, 2002) which we call the basic

skin color classiﬁer. The pixel-based skin color evi-

dence is then aggregated to classify the superpixels.

Finally, the CRF is applied to further improve the re-

sults. As CRF operates over superpixels, the compu-

tational overhead is minimal.

The proposed region-based technique achieved

91.44% true positive rate with 13.73% false positive

rate without CRF optimization and 91.17% true pos-

itive rate with 13.12% false positive rate with CRF

optimization. Grouping the pixel-based evidence into

superpixels increased the true positive rate by 1.44%

and reduced the false positive rate by 0.48%. More-

over, the region-based approach produced smoother

results than the pixel-based methods. Skin commonly

appears as regions of similar pixels, so treating skin as

a region is advantageous over treating it as an individ-

ual pixel. Due to the illumination, background reﬂec-

tion and other noise factors, pixel values vary greatly

and groupingthem into a region helps to removenoise

by collecting evidence from neighbouring pixels.

These results suggest that skin color detection

REGION-BASED SKIN COLOR DETECTION

305

should be region-based rather than pixel-based. Also,

by adding more constraints on the CRF similar to

(Shotton et al., 2006) , the detection rate can be im-

proved. Moreover, any better skin color classiﬁcation

method can be used as our basic skin color classiﬁ-

cation module and can be easily combined with our

region-based skin color detection framework deﬁned

in section 3 to improve the results.

REFERENCES

Boykov, Y. and Kolmogorov, V. (2004). An experi-

mental comparison of min-cut/max- ﬂow algorithms

for energy minimization in vision. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

26(9):1124–1137.

Boykov, Y., Veksler, O., and Zabih, R. (2001). Fast ap-

proximate energy minimization via graph cuts. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, pages 1222–1239.

Brown, D., Craw, I., and Lewthwaite, J. (2001). A som

based approach to skin detection with application in

real time systems. In Proceedings of the British Ma-

chine Vision Conference, volume 2, pages 491–500.

Fulkerson, B., Vedaldi, A., and Soatto, S. (2009). Class

segmentation and object localization with superpixel

neighborhoods. In Proceedings International Confer-

ence on Computer Vision, volume 5.

Jedynak, B., Zheng, H., and Daoudi, M. (2003). Maxi-

mum entropy models for skin detection. In Energy

Minimization Methods in Computer Vision and Pat-

tern Recognition, pages 180–193.

Jones, M. J. and Rehg, J. M. (2002). Statistical color mod-

els with application to skin detection. International

Journal of Computer Vision, 46(1):81–96.

Kakumanu, P., Makrogiannis, S., and Bourbakis, N. (2007).

A survey of skin-color modeling and detection meth-

ods. Pattern Recognition, 40(3):1106–1122.

Kawato, S. and Ohya, J. (2002). Automatic skin-color dis-

tribution extraction for face detection and tracking. In

International Conference on Signal Processing, vol-

ume 2, pages 1415–1418. IEEE.

Kolmogorov, V. and Zabih, R. (2004). What energy func-

tions can be minimized via graph cuts? IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

26(2):147–159.

Kruppa, H., Bauer, M., and Schiele, B. (2002). Skin patch

detection in real-world images. In Van Gool, L.,

editor, Pattern Recognition, volume 2449 of Lecture

Notes in Computer Science, pages 109–116. Springer

Berlin / Heidelberg.

Moore, A. P., Prince, S., Warrell, J., Mohammed, U., and

Jones, G. (2008). Superpixel lattices. In IEEE Con-

ference on Computer Vision and Pattern Recognition.

Pan, Z., Healey, G., Prasad, M., and Tromberg, B. (2003).

Face recognition in hyperspectral images. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 25(12):1552–1560.

Peer, P., Kovac, J., and Solina, F. (2003). Human skin colour

clustering for face detection. In International Confer-

ence on Computer as a Tool.

Phung, S. L., Chai, D., and Bouzerdoum, A. (2002). A uni-

versal and robust human skin color model using neural

networks. In Proceedings of International Joint Con-

ference on Neural Networks, volume 4, pages 2844–

2849.

Ren, X. and Malik, J. (2003). Learning a classiﬁcation

model for segmentation. In IEEE International Con-

ference on Computer Vision, volume 1.

Sebe, N., Cohen, I., Huang, T., and Gevers, T. (2004). Skin

detection: A bayesian network approach. In Proceed-

ings of the 17th International Conference on Pattern

Recognition, pages 903–906, Cambridge, UK.

Shotton, J., Winn, J., Rother, C., and Criminisi, A. (2006).

Textonboost: Joint appearance, shape and context

modeling for multi-class object recognition and seg-

mentation. Proceedings of European Conference on

Computer Vision, pages 1–15.

Soatto, S. (2009). Actionable information in vision. In Pro-

ceedings of the International Conference on Computer

Vision, volume 25.

Socolinsky, D. A., Selinger, A., and Neuheisel, J. D. (2003).

Face recognition with visible and thermal infrared im-

agery. Computer Vision and Image Understanding,

91(1-2):72–114.

Terrillon, J. C., Fukamachi, H., Akamatsu, S., and Shirazi,

M. N. (2000). Comparative performance of different

skin chrominance models and chrominance spaces for

the automatic detection of human faces in color im-

ages. In Fourth IEEE International Conference on

Automatic Face and Gesture Recognition, page 54.

Vedaldi, A. and Fulkerson, B. (2008). VLFeat: An open

and portable library of computer vision algorithms.

http://www.vlfeat.org.

Vedaldi, A. and Soatto, S. (2008). Quick shift and kernel

methods for mode seeking. Proceedings of European

Conference on Computer Vision, pages 705–718.

Wong, K. W., Lam, K. M., and Siu, W. C. (2003). A ro-

bust scheme for live detection of human faces in color

images. Signal Processing: Image Communication,

18(2):103–114.

Yang, M. H. and Ahuja, N. (1998). Detecting human faces

in color images. In International Conference on Image

Processing, 1998, volume 1, pages 127–130.

Zhu, Q., Cheng, K. T., Wu, C. T., and Wu, Y. L. (2004).

Adaptive learning of an accurate skin-color model.

In Sixth IEEE International Conference on Automatic

Face and Gesture Recognition, pages 37–42.

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

306