Perceptual color space represe... free essay sample

Perceptual color space represent color which is closely related with human vision. The main idea is to arrange all the colors by perceptual difference of color. In these color spaces-luminance (L) and chrominance is represented in XYZ coordinates. However, skin detection is not a physical property of an object in an image. It can be related as a perceptual concept of human vision. So, it is quite efficient to take a color space for skin detection that has the sensitivity of human vision. CIELAB and CIELUV are two examples of perceptually uniform color spaces. Perceptual uniformity defines that a small distortion to a component value can be equally observable across the range of values. Moreover, RGB and other well-known color spaces are not perceptually uniform. Non –linear transformation of perceptual color spaces corrects this shortcoming. But it needs complex transformation function for conversion. When an image is converted to CIE-LAB it may appear exactly as the previous color space. The reason is CIE-LAB and CIE-LUV contains all possible colors and since the translation of color is not related, so no deterioration in image quality. It may be little tricky since in usual color spaces (e.g., RGB, HSV) contains logical colors. But in LAB one of the channel contains no color and other two channels have dual color combinations without any contrast. The L channel is for lightness and there is no color value associated with it. It usually depicts the difference between presence of darkness and light. In addition to that, A is the color balance between green and Magenta and B is a balance between blue and yellow. For aforementioned advantages Zarit et al. [47], Yang and Ahuja [25], Schumeyer and Barner [48] used this color space. However, I also consider this color space for clustering skin pixels from an image.Color space comparisonsColor spaces obviously affect the performance of skin detectors. Several authors experimented different results for the influence of color space choice on skin segmentation algorithms performance. In [49], author claimed that TSL is the best color space for skin detection While combined with Gaussian and Mixture of Gaussian models. Comparing the performance of variety of color spaces on single Gaussian model has been yielded to YCbCr for the best performance [50]. On the other hand, Montenegro et al. compared RGB, HSV, YCbCr, CIE-Lab and CIE-Luv implementing SFA dataset. Matthews correlation coefficient (MCC) was used as a performance metric and Gaussian classifier as the method. CIE-Lab outperforms all others and normalized RGB performs the least [51]. Furthermore, performance of RGB based techniques suffers in decimating skin-pixels while objects increases [52]. A comparison study is done using Gaussian and histogram approaches for a dataset of 805 color images [53]. This works claimed that the choice of color space significantly changes the performance, and HSI color space accompanying with histogram model outperform others. In comparison of three different color spaces (HIS, RGB, CIE-Lab and YCbCr), in terms of classification error is obtained in HIS and YCbCr models [7]. A recent study [54] claims YPbPr outperforms other models. Besides Gaussian approaches, Gonzà ¡lez et al. [55] compared the performance of 10 common color spaces based on k-means clustering algorithm and conclude that HSV, YCgCr, and YDbDr are the best color spaces. Nalepa et al. [56] showed that statistically combination of RGB and HSV outperforms other color spaces. In other hand, artificial neural network has been subjected for skin segmentation in variety of color spaces yielding to the conclusion that YIQ is the best choice [57]. The idea to adapt an optimal chrominance color space rather than a segmentation model has been suggest in [58]. Here a non-liner transformation between YUV and new TSL* color space is used to boost the segmentation. In summary, performance of color space based skin detection is highly depend on various factors including used methods. To make a fair comparison, all influencing factors should be considered. As observed, different authors take account of different training, testing and validation data. As a result, optimum color space for skin segmentation has changed to adapting optimal skin detection models. Besides, in neural networks, the performance of classifier is directly depend on the number of neurons, as well as the initial guess of the weights. 2.2 Skin Detection MethodsSkin detection processes may be categorized into diverse categories which are not mutually distinct. Statistical strategies are based totally on records extracted from histogram of training skin and non-skin pixels. Non-parametric or parametric models are developed with the aim of acquiring the probability that a pixel belongs to skin or non-skin class. Artificial neural networks (ANN) are very beneficial gear in both estimation of skin distribution or direct classification of the pixels. Adaptive techniques often reach better accuracies with the cost of computation. SVM based techniques also are used for classification of skin pixels in applications. 2.2.1 Explicitly Defined Boundary ModelsThe main difference in skin color is in its intensity (brightness) rather than the chrominance (color) [59, 55, 57]. Those techniques have a pixel based processing scheme in which for any given pixel, methods are investigated to determine the class of that pixel. Theyre very prominent particularly due to their simple and quick training, low cost implementation and fast processing. However, several parameters are concerned in degrading the performance of classifiers, consisting of their static nature, excessive dependence on training images, effectiveness of rules and inability to cope with maximum skin detection challenges. Kovac et al. [60] proposed a way of explicitly defined boundary model using RGB color space in two of daylight and flashlight situations which has been reutilized in [61,62,63,64]. Orthogonal color areas are regularly utilized in case of explicitly described boundary models. Sagheer et al. [65] exploited cases of regular lights and distinctive lighting fixtures. In order to detect faces, in [66, 67], rules are dynamically reconfigured based on pixels value. Zahir et al. [68] proposed an easy boundary version using HSV color space for indoor and outdoor conditions. Combination of color spaces has been also effective to reinforce the performance of explicitly defined boundary methods. Thakur et al. [69] employed RGB, CbCr and HSV color method. Rules are set independently for triple color spaces and outcomes are clearly fused for taking final selection. Furthermore, fusion of the result of different color space to reduce false positives has been a common procedure; some examples are: RGB and YCbCr [70], YCbCr and YUV [8], HSV and YUV [71], HSV and YCbCr [72,73], RGB and YUV [74] and HSV and YCgCr [75].2.2. 2 Statistical ModelsSkin detection is a probabilistic problem and lots of techniques based on general distribution of skin and skin color, every in a selected color model, were developed. Based totally on an extensive training set, these strategies estimate the probability that a located pixel is associated with skin. 2.2.2.1 Non-parametric (histogram based) modelsThere is no any specific definition of probability density characteristic in non-parametric techniques. Single histogram primarily based Look-UP-Table (LUT) model is a common approach in modeling skin color cluster. In this technique, by using a set of training skin pixels, distribution of skin pixels in a selected color space is received. The training technique is simple but populating the histogram calls for a large skin dataset. But, Bayesian classifier considers two histograms of skin and non-skin pixels. In [76, 77], a Bayesian classifier based on YCbCr color space is employed. Erdem et al. [61] employed Bayesian based post filtering technique to reduce false positives of Viola-Johns face detector. Zarit et al. [78] performed a comparison study of overall performance of histogram method in Fleck HS, HSV, RGB, CIE Lab and YCbCr. Phung et al. [7] examined the tradeoff among the wide variety of bins in line with channel and detection rate.There are numerous gain and drawbacks associated with non-parametric techniques. The benefit includes- training and implementation, rapid training, independency at the shape of cluster [5] are the primary blessings. However, one main drawback of these strategies is their dependency at the training set which requires gathering of a massive range of skin pixels. Further, large memories are required to implement histogram based models especially whilst fine resolutions are required. 2.2.2.2 Parametric modelsParametric method which include single Gaussian method (SGMs), Gaussian mixture models (GMMs), cluster of Gaussian models (CGMs), Elliptical models (EMs), etc, are evolved to compensate LUT shortcomings consisting of excessive storage requirement. Similarly, they generalize thoroughly with a tremendously smaller amount of training set. In SGM, there should be a smooth Gaussian distribution around the mean vector. GMM is advanced to version more complex distributions as a normalized weighted sum of Gaussian PDFs [25]. It compensates the inability of SGM in managing out of control conditions in a general skin segmentation problem, similarly to the truth that SGM is not capable of approximating the actual distribution due to the asymmetry of distribution in its peak [79]. The most appealing features of GMM models are their easy evaluation procedure and occasional memory cost. The training technique is but, longer than the former methods. A comparative look at the overall performance of single and combination of Gaussian distributions in [80] confirmed that mixture method improve the performance only in a relevant operating area.Several statistical skin detection strategies have been elucidated in this subsection; all classified in non-parametric and parametric processes. Non-parametric methods are primarily based on histogram of skin and non-skin pixels in a predefined training set. But, in compare with parametric