Shading Correction in Human Skin Color ImagesPablo G. Cavalcanti and Jacob Scharcanski

Instituto de InformaticaUFRGS, Brazil

1 Introduction

In the context of computer vision, there are several applications that require human skin image processing andanalysis. However, the color of human skin structures can be distorted by shading effects. The occurrence ofshading effects depends on the color of the object and the lighting conditions. Nevertheless, other factors also caninfluence the surface color captured by a camera, like the surface roughness, the relative position of the reflectiveskin area with respect to the light sources and the camera (Shapiro & Stockman, 2001). Specifically, humanskin images are affected by these factors and the analysis of these images can become even more difficult if theillumination is uneven, unless it is correctly modeled and corrected, as we discuss next.

Man-machine interaction is an important human skin color imaging application. Often, color imagescontaining human skin are used in head pose estimation or in face recognition systems, and shading effectsmay distort some important facial features in the captured images (e.g., eyes, nose, head geometry). Severalefforts have been directed towards the development of hand gesture recognition systems. Also, in hand gesturerecognition applications usually often it is not feasible to control the illumination during image acquisition due tothe complex hand motion, and shading may occlude parts of the hand and make the scene analysis more difficult.In these cases, an automatic preprocessing step to mitigate the impact of the shading effects may help increasethe efficiency of these systems, as we will illustrate later.

Nowadays, computer vision is an important ally in medical problems. For example, computer-aided sys-tem have been developed in the recent last years to enable remote surgeries or help physicians in image-baseddiagnosis. An area that has been receiving a lot of attention is dermatology, probably because it is the most visualspecialty in medicine. In teledermatology, often a color image of a skin lesion acquired with a standard camerais transmitted to a specialist (or to a pre-screening system), without paying special attention to the illuminationconditions (Whited, 2006)(Massone et al., 2008). Nevertheless, the illumination may influence significantly thequality of the lesion visualization, and impact on the physician diagnosis, as well as it may limit the efficiencyof the pre-screening system. If the illumination condition is inadequate, in general we obtain low diagnosis ac-curacy of pigmented skin lesions. These lesions usually are darker than healthy skin, and automatic approachesto segment such lesions tend to confuse shading areas with lesion areas. As a consequence, the early detectionof malignant cases is more difficult without removing shading effects from the input images. Considering thatmelanoma is the most dangerous type of pigmented skin lesion, and that this disease results in about 132000 casesglobally each year (World Health Organization, 2011), any contribution to improve the quality of these imagescan be an important step to increase the efficiency of teledermatology and pre-screening systems, and to help todetect cases in their early-stages.

In the following sections, we discuss the automatic shading effect attenuation in human skin images. InSection 2, we describe a method for skin image shading attenuation, and discuss the experimental results inSection 3. We emphasize the benefits of shading attenuation for the color image analysis of face, hand gestureand pigmented skin lesions images. We compare this methodology with other techniques available in the literaturein Section 4. Finally, in Section 5 we present our conclusions.

2 Shading Attenuation in Human Skin Color Images

The approach for shading attenuation discussed in the Section improves on the method proposed by Soille (Soille,1999). The method in (Soille, 1999) corrects uneven illumination in monochromatic images with a simple oper-ation:

R(x,y) = I(x,y) / M(x,y), (1)

where, R is the resultant image, I is the original image, M = I • s is the morphological closing of I by thestructuring element s, and (x,y) represents a pixel in these images. The main idea behind Soille method is to usethe closing operator to estimate the local illumination, and then correct the illumination variation by normalizingthe original image I by the local illumination estimate M. The division in Eq. 1 relights unevenly illuminatedareas, without affecting the original image characteristics. Unfortunately, it is often difficult to determine anefficient structuring element for a given image, specially for human skin images that have so many distinctfeatures, such as hair, freckles, face structures, etc. In this way, the results tends to be unsatisfactory for thistype of images, as can be seen in Figs. 1(b)-(c).

(a) (b) (c)

(d) (e) (f)

Figure 1: Shading attenuation in a pigmented skin lesion image : (a) Input image; (b) Morphologicalclosing of Value channel by a disk (radius = 30 pixels); (c) Unsatisfactory shading attenuation afterreplacing the Value channel by R(x,y), as suggested by Soille (Soille, 1999); (d) Local illuminationbased on the obtained quadric function; (e) 3D plot of the obtained quadric function; (f) Shadingattenuation by using our approach.

The presented method modifies the Soille approach by providing a better local illumination estimate M.In order to provide this local illumination estimate, it starts by converting the input image from the original RGBcolor space to the HSV color space, and then retaining the Value channel V . This channel presents a highervisibility of the shading effects, as observed originally by Soille.

This approach is inspired on the computation of shape from shading (Shapiro & Stockman, 2001). Thehuman body is assumed to be constituted by curved surfaces (e.g. arms, back, faces, etc.) and, in the sameway humans see, digital images present a smoothly darkening surface as one that is turning away from the viewdirection. However, instead of using this illumination variation to model the surface shape, this information isused to relight the image itself.

Let S be a set of known skin pixels (more details in Section 3). This pixel set is used to adjust the followingquadric function z(x,y):

z(x,y) = P1x2 +P2y2 +P3xy+P4x+P5y+P6, (2)

where the six quadric function parameters Pi (i = 1, ...,6) are chosen to minimize the error ε:

ε =N

∑j=1

[V (S j,x,S j,y)− z(S j,x,S j,y)]2 , (3)

where, N is the number of pixels in the set S, and S j,x and S j,y are the x and y coordinates of the jth element of theset S, respectively.

Calculating the quadric function z(x,y) for each image spatial location (x,y), an estimation z of the localillumination intensity in the image V is generated. Replacing M(x,y) by z(x,y), and I(x,y) by V (x,y) in Eq. 1,we obtain the image R(x,y) normalized with respect to the local illumination estimate z(x,y). The final step is toreplace the original Value channel by this new Value channel, and convert the image from the HSV color spaceto the original RGB color space. As a consequence of this image relighting, the shading effects are significantlyattenuated in the color image. Figs. 1(d)-(e) illustrate the results obtained with this shading attenuation method.

3 Shading Attenuation: Experimental Results and Discussion

As mentioned in Section 2, the shading attenuation approach discussed in this chapter is initialized by a set ofpixels S known to be associated with healthy skin areas. In this section, we discuss how to select this set ofpixels S in three typical applications of human skin color image analysis (i.e. image segmentation), namely, thesegmentation of faces, hands and pigmented skin lesions in color images. Our goal is to show that our shadingattenuation approach helps in the image analysis in these applications, making the processing steps simpler.

3.1 Face Segmentation in Color Images

Face segmentation is a common need in man-machine interaction applications. In fact, face segmentation canbe used in many applications, like face recognition, speech understanding, head pose/motion detection (likehead shaking, eye sight direction, etc.), or active speaker detection. Usually such applications require accurateface segmentation for an adequate system usability. Also, the detection of face expression involves extractingsensitive features from facial landmarks such as the regions surrounding the mouth, nose, and eyes of a normalizedimage (Mitra & Acharya, 2007). So, faces usually are localized and segmented as the initial step to approach anyof the above mentioned applications, but that may not be an easy task under uneven illumination conditions.

In addition to the shading effects, in this case we must be aware that a face may be located at virtuallyany image location. In this discussion, face segmentation will use color information only, i.e. we will search for

image locations with colors similar to human skin tones. According to previous research (Vassili et al., 2003), apixel can be associated to a skin region if :

R > 95 ∧ G > 40 ∧ B > 20 ∧ (4)

max(R,G,B)−min(R,G,B)> 15 ∧|R−G|> 15 ∧ R > G ∧ R > B,

where, ∧ denotes the logical operator and. However, as can be seen in Fig. 2, this method is not able to identify allthe skin pixels correctly. Although this criterion to determine pixels associated to skin color is used often (Vassiliet al., 2003), it can be very imprecise in practical situations, specially where there is image shading. On otherhand, this method can locate with some accuracy some skin regions in images, and these pixels can be used asthe set S of known skin pixels. We adopt this method to find the set S, since all we need is a set of adjacent imagepixels with skin color (i.e. likely to be located in skin regions) to initialize our error minimization operation (seeEqs. 2 and 3), and erroneously located pixels should not influence significantly the final result.

(a) (b) (c)

Figure 2: Illustration of skin pixels localization using Eq. 4 : (a) Input image; (b) Binary mask; and(c) adjacent pixels identified as human skin.

Given the set S by using Eq. 4, the shading effects in the face image can be attenuated. To demonstrate theefficacy of our method in this application, we show the face segmentations with, and without, shading attenuationusing a known Bayes Classifier for the pixels based on their corrected colors (Vassili et al., 2003). A pixel isconsidered skin if:

P(c|skin)P(c|¬skin)

> θ, (5)

where θ = κ× 1−P(skin)P(skin)

. (6)

In Eq. 5, the a priori probability P(skin) is set to 0.5, since we use the same number of samples for eachclass (i.e. 12800 skin pixels and 12800 non-skin pixels). The constant κ also is set to 0.5, increasing the chanceof a pixel be classified as skin, and P(c|skin) and P(c|¬skin) are modeled by Gaussian joint probability densityfunctions, defined as:

P =1

2π|∑ |1/2 × e−12 (c−µ)T

∑−1(c−µ), (7)

where, c is the color vector of the tested pixel, and µ and ∑ are the distribution parameters (i.e., the mean vectorand covariance matrix, respectively) estimated based on the training set for each class (skin and non-skin).

Figure 3: Face segmentation examples. In the first and second columns the original images are shown,as well as their respective segmentation results. In the third and fourth columns, the images after theapplication of our shading attenuation method are shown, and their respective segmentation results.

Figs. 3 and 4 illustrate some face segmentation examples. These face images are publicly available in thePointing’04 dataset (Gourier et al., 2004). The images in Fig. 3 show four different persons, with different phys-ical characteristics and different poses (i.e. angles between their view direction and the light source), resultingin different shading effects. Clearly, the skin pixels, and consequently the faces, are better segmented after weapply our shading attenuation method in all these different situations. In Fig. 4, we present four examples of thesame person but with different head poses (the angle between her view direction and the light source). It shallbe observed that even when the face is evenly illuminated, the face is better segmented after using our shadingattenuation method. However, inaccuracies may occur near the facial features partially occluded by cast shadows(e.g. near the nose and the chin). Based on these results, it should be expected that algorithms that extract fa-cial features (e.g., eyes, mouth and nose) would perform their tasks more effectively, which helps man-machineinteraction systems.

Figure 4: Face segmentation examples for the same person with different head poses. In the first andsecond columns are shown the original images, and their respective segmentation results. In the thirdand fourth columns, are shown these images after applying our shading attenuation method, and therespective segmentation results.

3.2 Hand Segmentation in Color Images

In the same way that face segmentation is important in man-machine interaction applications, sometimes inter-action requires a precise hand segmentation (e.g., when hand gestures are used in the man-machine interaction).Hand gestures can be used to facilitate the access for people with difficulty in using traditional input devices (e.g.,keyboards), or make the man-machine interaction easier and more natural.

Usually, hand gestures image interpretation require dynamic information, and the hand and/or arm motionmust be segmented to capture temporal segmentation. The automatic segmentation of the gesture start and endpoints in time and in space are important in this application. Sometimes, gesture detection is affected by thecontext (e.g. preceding and subsequent gestures) (Mitra & Acharya, 2007), and gesture recognition often isapproached using statistical models like particle filtering and Hidden Markov Models. However, before usingsuch techniques for tracking hands and/or recognizing gestures, a frame-by-frame hand segmentation is needed.So, the image area corresponding to a hand is localized first, and then features are extracted for tracking andrecognition.

(a) (b) (c)

(d) (e) (f)

Figure 5: Illustration of skin pixels localization in typical hand gestures images using Eq. 8 : in thefirst row, the input images; in the second row, the adjacent pixels identified as human skin.

Hand segmentation is a problem very similar to face segmentation, since both are based on skin-pixellocalization, and different classification techniques such as thresholding, Gaussian classifier, and multilayer per-ceptrons could be used for this task. Recently, Dardas and Georganas (Dardas & Georganas, 2011) suggestedusing the HSV color space, and based on this approach a pixel can be associated to a hand skin region if :

0◦ < H < 20◦ ∧ 75 < S < 190, (8)

where, ∧ denotes the logical operator and. However, as can be observed in Figure 5, this thresholding techniquehas the similar problems to the RGB based thresholding technique that we presented in Section 3.1 (see Eq. 4).Although it localizes the skin area with some accuracy, it is very susceptible to noise and may fail specially inpixels affected by shading effects.

Nevertheless, the Bayes Classifier could be used to improve the segmentation results as follows. The handimages can be processed with a similar algorithm to that presented in Section 3.1. In other words, the shadingeffects are modeled and attenuated using an RGB-based thresholding technique, two Gaussian joint probabilitydensity functions P(c|skin) and P(c|¬skin) are modeled, and a pixel is defined as a skin-pixel or backgroundaccording to Eq. 5.

We present some illustrative examples of shading attenuation in Fig. 6 based on the hand images of theSebastien Marcel database (Marcel, 1999), which often is used as a benchmark in hand gesture recognition. Thereader can observe that the areas affected by shading are relighted. As can be seen in Fig. 7, the Bayes Classifierachieves more accurate results than the thresholding method proposed by Dardas and Georganas before described.Moreover, a better segmentation result is obtained if shading attenuation is used as a pre-processing stage. Animproved hand segmentation tends to increase the chance of correct hand gesture recognition, since the extractedhand features are more reliable in this case.

(a) (b) (c) (d) (e) (f)

(g) (h) (i) (j) (k) (l)

Figure 6: Illustration of the application of our shading attenuation approach to hand gestures images:the input images are in the first row; after shading attenuation, the images are shown in the second row.

3.3 Pigmented Skin Lesion Segmentation in Color Images

Several methods have been proposed for analyzing pigmented skin lesions in dermoscopy images (Maglogiannis& Doukas, 2009). Dermoscopy, is a non-invasive technique that magnifies submacroscopic structures with thehelp of an optical lens (a dermoscope) and liquid immersion. Usually, the first step of a computer-aided diagnosissystem for pigmented skin lesions is to segment the lesion areas, discriminating lesion and healthy skin pixels.Many segmentation methods have also been proposed for this task (Celebi et al., 2009). However, the dermoscopeusually has a light source so the captured images are not affected by illumination artifacts. However, dermoscopesare tools used by experts to help in the diagnosis, but there are practical situations where a non-specialist wishesto have a qualified opinion about a suspect pigmented skin lesion, but only standard camera imaging is availableon site (i.e., telemedicine applications).

Our discussion in this section is based on standard camera images, i.e. standard photographs of pigmentedskin lesions. It is not trivial to acquire reliable standard photographs of these lesions, since pigmented skin lesionsusually are only a few millimeters width, the camera should be placed near the the skin and the acquired imageusually is affected by shading effects. Considering the telemedicine context, the physician who receives suchan image probably would have difficulties in screening it. In the same way, pre-screening systems also couldhave difficulties to automatically pre-screen such a lesion image. Next, we discuss how shading attenuationcould be helpful in this situation, justifying its use in telemedicine and in standard camera imaging. To illustratethe effectiveness of our shading attenuation approach, we compare the segmentation results for pigmented skinlesions with and without the application of our method.

Let us focus on the image skin area containing the lesion. During image acquisition, the lesion is capturedas the central portion of the image, and is surrounded by healthy skin. Therefore, we assume the four imagecorners to contain healthy skin. This assumption is common in dermatological imaging, and has been adoptedby researchers in this field (Celebi et al., 2008) (Melli et al., 2006). Therefore, we use 20 × 20 pixel sets aroundeach image corner, and determine S as the union of these 1600 pixels (i.e. the four pixel sets).

Unfortunately, segmentation methods for pigmented skin lesions on standard camera images did not re-ceive much attention in the literature as the segmentation methods for pigmented skin lesions for dermoscopyimages. Usually, thresholding-based techniques are used in the segmentation, like the Otsu’s Thresholding

Figure 7: Hand segmentation examples. In the first and second columns are shown the original images,and their respective segmentation results. In the third and fourth columns, are shown images after theapplication of our shading attenuation method and their segmentation results, respectively.

method (Otsu, 1979), which has been widely used in grayscale images (Manousaki et al., 2006; Ruiz et al.,2008; Tabatabaie et al., 2009). Furthermore, Cavalcanti et al. (Cavalcanti et al., 2010) also employed this thresh-olding scheme to the Red channel (R of the RGB color space), trying to take advantage of the fact that healthyskin usually has a reddish tone. This method assumes two pixel classes, namely healthy and unhealthy skin pixels,and searches exhaustively for the threshold th that minimizes the total intra-class variance σ2

w(th), defined as the

weighted sum of variances of the two classes:

σ2w(th) = ω1(th)σ2

1(th)+ω2(th)σ22(th), (9)

where ωi are the a priori probabilities of the two classes separated by the threshold th, and σ2i are their intra-class

variances. Minimizing the intra-class variance is equivalent to maximizing the inter-class variance σ2b(th):

σ2b(th) = σ

2−σ2w(th) (10)

= ω1(th)ω2(th) [µ1(th)−µ2(th)]2 ,

where σ2 is the image pixels variance, and µi are the class means. Computed the th threshold, the lesion pixelscorrespond to the pixels with values lower than th.

Following the Otsu’s method usually a post-processing step is employed, and often is constituted by suc-cessive morphological operations, to eliminate regions associate to artifacts that may be thresholded (besides theskin lesion). A thorough discussion of this topic is out of the scope of this work, and we suggest the reading someadditional literature (Manousaki et al., 2006; Ruiz et al., 2008; Tabatabaie et al., 2009; Cavalcanti et al., 2010)for more details.

Exploring a different line of work, Alcon et al. suggested that Otsu’s method may over-segment the lesionarea. So, they proposed a different thresholding method specific for pigmented skin lesions acquired with standardcamera images. They observed that, although the lesion intensities distribution fl(x) is unknown, the distributionfs(x) of the skin corresponds to a Gaussian-like distribution:

fs(x) = A e−(x−µs)2

2σ2s , (11)

where, µs is the mean value of healthy skin pixel intensities. Being fl+s the distribution of grayscale intensities ofthe whole image, µs is determined by the corresponding intensity value of the highest peak of fl+s. Since fl+s =fl + fs, and µl (the mean value of lesion pixels) always is lower than µs, this distribution can be approximated as:

fl+s(x) ={

fs(x) , x≥ µs

fl(x)+ fs(x) , x < µs. (12)

Therefore, based on this assumption, the skin pixels distribution can be estimated as :

fs(x) ={

fl+s(2µs− x) , x < µs

fl+s(x) , x≥ µs, (13)

and, consequently, the lesion pixels distribution can be estimated as :

fl(x) = fl+s(x)− fs(x). (14)

Finally, the means E(Xs) and E(Xl) of the distributions fs(x) and fl(x), respectively, are used for thecomputation of the threshold T as follows :

T =E(Xs)+ E(Xl)

2, (15)

and, as in the Otsu’s method, the pixels with values lower than the computed threshold, are segmented as lesionpixels.

(a) (b)

Figure 8: Pigmented skin lesion image segmentation using thresholding-based methods. (a) The inputimage converted to grayscale (after pre-processed by the shading attenuation method). (b) The plot ofhistogram of figure (a), and the Otsu and Alcon thresholds obtained.

Figure 8 presents the difference between Otsu’s and Alcon et al. thresholds. The reader may observe thateach technique determines a different threshold value to separate the lesion pixels (pixels intensities above thethreshold value) and the healthy skin pixels (pixels intensities higher then the threshold value).

The segmentation results are presented in Figure 9. Malignant pigmented skin lesions (melanomas) fromthe Dermnet Skin Disease Image Atlas (Dermnet Skin Disease Image Atlas, 2010) were processed with the shad-ing attenuation method, and then segmented by the three segmentation techniques described previously (i.e.,Otsu’s Thresholding method applied on grayscale images, Otsu’s Thresholding method applied on the Red chan-nel, and the Alcon et al. Thresholding method). The results obtained without the application of the shadingattenuation step are also presented for the sake of comparison. In similar way, Figure 10 also presents segmenta-tion results, but for benign lesions (dysplastic nevi).

As can be seen, the obtained segmentation results are better if shading attenuation is used as a pre-processing step. Considering that the lesion shape and boundary provide crucial information for discriminat-ing malignant and benign lesions, this improved segmentation results potentially could be very helpful for pre-screening skin lesions.

It is important to observe that our shading attenuation approach may fail in some situations, as illustratedin Fig. 11. The typical situations illustrated in Fig. 11 are: (a) our method is adequate to model and attenuate theglobal illumination variation (which changes slowly), but tends to have limited effect on local cast shadows; and(b) our approach tends to fail on surface shapes that are not locally smooth, since the quadric function is not ableto capture the local illumination variation in this case. In such cases, the segmentation method may cause healthyand unhealthy skin areas to be confused. Possibly, better results could be achieved in such cases by acquiring theimages in a way that surface shapes are smoother and illumination varies slowly across the scene.

Figure 9: Malignant pigmented skin lesions segmentation examples. From the first to the fourth row,results without the application of the shading attenuation method. From the fifth to the eighth row,results after using the shading attenuation method as a color image preprocessing step. In the firstcolumn, the original color images are shown; in the second column, results of the Otsu’s Thresholdingmethod applied on grayscale images are shown; in the third column, results of the Otsu’s Threshold-ing method applied on the Red channel are shown; in the fourth column, results of the Alcon et al.Thresholding method are shown.

Figure 10: Benign pigmented skin lesions segmentation examples. From the first to the fourth row, re-sults without the application of the shading attenuation method. From the fifth to the eighth row, resultsafter using the shading attenuation method as a color image preprocessing step. In the first column,the original color images are shown; in the second column, results of the Otsu’s Thresholding methodapplied on grayscale images are shown; in the third column, results of the Otsu’s Thresholding methodapplied on the Red channel are shown; in the fourth column, results of the Alcon et al. Thresholdingmethod are shown.

Figure 11: Illustrations of cases where our shading attenuation method tends to fail, such as castshadows (first line) and surface shapes not well modeled by quadric functions (second line). The firstand second columns show the original images and their respective segmentation results (using Otsu’sThresholding Method). The third and fourth columns show the resulting images after the applicationof our shading attenuation method, and the respective segmentation results (using Otsu’s ThresholdingMethod).

4 Comparison With Other Shading Attenuation Methods

It is important to compare the results of the proposed shading attenuation method with other techniques frequentlyused in the literature, such as homomorphic filtering (Petrou & Petrou, 2010) and the Retinex model (Jobsonet al., 1997). We already presented in Section 2 the method proposed by Soille (Soille, 1999), which inspired themethodology presented in this chapter, and compared results.

Homomorphic filtering relies on a non-linear transform (i.e., logarithm) of the image intensities, and theillumination component can be removed by suppressing the low-frequency image components. The remaininghigher-frequency components are associated with the surfaces reflectance showing in the image (Petrou & Petrou,2010). However, selecting the appropriate high-pass filtering can be challenging, given the different character-istics of surfaces (e.g. human skin images). Moreover, the elimination of the low-frequency components mayalter the image characteristic making the image segmentation more difficult, since edges and details are sharp-ened (Petrou & Petrou, 2010). In Fig. 12 we present examples of applying the homomorphic filtering to the Valuechannel of our tested input images.

The Retinex model has been proposed to achieve lightness-color constancy, i.e. preserve the perceivedcolor of objects relatively constant under varying illumination conditions (Jobson et al., 1997). Although theRetinex may correct the illumination condition, it requires parameters specification and may negatively affectthe image colors and contrast. In Fig. 12, we present examples of applying the Retinex model implementationproposed by Jobson et al. (Jobson et al., 1997).

Besides the results obtained with the two methods discussed above, we show in Fig. 12 the results ob-tained with the shading attenuation method presented in this chapter. As can be seen, the proposed method tendsto attenuate shading better then the other methods used in our comparison, while enhancing the skin-backgroundcontrast in human skin images. We also superimposed the boundaries of the segmentation results obtained withthe methods presented in Section 3) and show that our shading attenuation method tends to help in image seg-mentation. Our shading attenuation method has been proposed specifically for this type of images, and it also hasthe advantage of being fully automatic, since it does not require parameters tuning and is adaptive to the imagedata.

Figure 12: Comparison with other methods. In the first column, the original images. In the secondcolumn, after homomorphic filtering. In the third column, after applying the Retinex model. In thefourth column, after applying the shading attenuation method presented in Section 2. The red curvesindicate the region borders after segmenting these images.

5 Conclusions

We discussed an approach for attenuating the shading effects in human skin images. According to the proposedapproach, given a set of pixels known to be skin, a quadric function is adjusted to this pixel set to derive a modelto relight all the skin area in the image.

A set of experiments are used to illustrate how the proposed approach could be applied to some typicalcolor image analysis problems where human skin imaging is of central importance. It has been demonstratedhow to automatically determine the set of known skin pixels, and how the application of the proposed methodmay increase the segmentation accuracy, and potentially contribute to improve the overall system efficiency.Considering the man-machine interaction problem, our experiments suggest that the shading attenuation methodcan improve the robustness of face and hand gesture recognition. We also showed that in the case of pigmentedskin lesion segmentation, shading attenuation method helps improving the lesion detection, and, hopefully, cancontribute for the early identification of skin cancer cases.

We also shall observe that our approach may not improve the image quality in situations where we havecast shadows or surface shapes that are not locally smooth. Possibly, the use of more complex quadric functionscould help solve these difficulties with our approach.

References

Cavalcanti, P., Yari, Y., & Scharcanski, J. (2010). Pigmented skin lesion segmentation on macroscopic images. In Proceedings ofthe 25th International Conference on Image and Vision Computing New Zealand.

Celebi, M., Iyatomi, H., Schaefer, G., & Stoecker, W. V. (2009). Lesion border detection in dermoscopy images. ComputerizedMedical Imaging and Graphics, 33(2), 148 – 153.

Celebi, M. E., Kingravi, H. A., Iyatomi, H., Aslandogan, Y. A., Stoecker, W. V., Moss, R. H., Malters, J. M., Grichnik, J. M.,Marghoob, A. A., Rabinovitz, H. S., & Menzies, S. W. (2008). Border detection in dermoscopy images using statistical regionmerging. Skin Res. Technol., 14(3), 347–353.

Dardas, N. & Georganas, N. (2011). Real-time hand gesture detection and recognition using bag-of-features and support vectormachine techniques. Instrumentation and Measurement, IEEE Transactions on, 60(11), 3592 –3607.

Dermnet Skin Disease Image Atlas (2010). http://www.dermnet.com.

Gourier, N., Hall, D., & Crowley, J. L. (2004). Estimating face orientation from robust detection of salient facial features. InProceedings of Pointing 2004, ICPR, International Workshop on Visual Observation of Deictic Gestures Cambridge, UK.

Jobson, D., Rahman, Z., & Woodell, G. (1997). Properties and performance of a center/surround retinex. Image Processing, IEEETransactions on, 6(3), 451 –462.

Maglogiannis, I. & Doukas, C. (2009). Overview of advanced computer vision systems for skin lesions characterization. InformationTechnology in Biomedicine, IEEE Transactions on, 13(5), 721 –733.

Manousaki, A. G., Manios, A. G., Tsompanaki, E. I., Panayiotides, J. G., Tsiftsis, D. D., Kostaki, A. K., & Tosca, A. D. (2006). Asimple digital image processing system to aid in melanoma diagnosis in an everyday melanocytic skin lesion unit: a preliminaryreport. International Journal of Dermatology, 45(4), 402–410.

Marcel, S. (1999). Hand posture recognition in a body-face centered space. In CHI ’99 extended abstracts on Human factors incomputing systems, CHI EA ’99 (pp. 302–303). New York, NY, USA: ACM.

Massone, C., Wurm, E. M. T., Hofmann-Wellenhof, R., & Soyer, H. P. (2008). Teledermatology: an update. Semin. Cutan. Med.Surg., 27(1), 101–105.

Melli, R., Grana, C., & Cucchiara, R. (2006). Comparison of color clustering algorithms for segmentation of dermatological images.In J. M. Reinhardt & J. P. W. Pluim (Eds.), Medical Imaging 2006: Image Processing, volume 6144 (pp. 61443S).: SPIE.

Mitra, S. & Acharya, T. (2007). Gesture recognition: A survey. Systems, Man, and Cybernetics, Part C: Applications and Reviews,IEEE Transactions on, 37(3), 311 –324.

Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man and Cybernetics,9(1), 62–66.

Petrou, M. & Petrou, C. (2010). Image Processing: The Fundamentals. John Wiley & Sons, 2nd. edition.

Ruiz, D., Berenguer, V. J., Soriano, A., & Martin, J. (2008). A cooperative approach for the diagnosis of the melanoma. InProceedings of the 30th. Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),volume 2008 (pp. 5144–5147).

Shapiro, L. & Stockman, G. (2001). Computer Vision. Prentice Hall.

Soille, P. (1999). Morphological operators. In B. Jahne, H. Haußecker, & P. Geißler (Eds.), Handbook of Computer Vision andApplications, volume 2 chapter 21, (pp. 627–682). San Diego: Academic Press.

Tabatabaie, K., Esteki, A., & Toossi, P. (2009). Extraction of skin lesion texture features based on independent component analysis.Skin Research and Technology, 15(4), 433–439.

Vassili, V. V., Sazonov, V., & Andreeva, A. (2003). A survey on pixel-based skin color detection techniques. In Proc. Graphicon-2003 (pp. 85–92).

Whited, J. D. (2006). Teledermatology research review. Int. J. Dermatol., 45(3), 220–229.

World Health Organization (2011). How commom is skin cancer? http://www.who.int/uv/faq/skincancer/en/index1.html.