Skin detection free essay sample

IntroductionSkin detection is perhaps the most widely used primitive in human image processing research. Skin detection mostly used as a primary step in various human concerned image processing applications. Skin detection is method of discriminating human skin pixel from non-skin pixels in an image or video [1]. It is one of the prominent research area in human computer interaction, face detection, face tracking [2, 3], gesture recognition [4], computational health informatics, web content filtering and many more. Skin detection is used as a cue for detecting people in real life images. The main challenge is to make skin detection robust to the large variations in appearance that can occur. However, there are various factors that make skin detection challenging. Among them variations in illumination, various ethnicity people with many skin tones, presence or absence of shadows in an image or videos, various background color and objects including wood, cloths and their similarity to skin, human hair with different variations and their resemblance to skin color, using makeup that changes the natural skin color and different camera characteristics make skin detection problem hard. We will write a custom essay sample on Skin detection or any similar topic specifically for you Do Not WasteYour Time HIRE WRITER Only 13.90 / page Efficient handling of aforementioned challenges demands a model that is capable of differentiate skin and non-skin pixels. But until now that seems not to be achieved. In this thesis, a skin detection model is proposed which can overcome the challenges and perform better in real world skin detection problem.Researchers have been working tirelessly to find a technique which will be able to detect skin in spite of present challenges. However, Skin detection problem can be considered a binary classification problem, meaning, a pixel can be classified whether it is a skin pixel or not. There are mainly two types of skin detection, either pixel based or region based. In pixel based detection, a pixel is classified compared with its neighbor either as a skin pixel or not. Skin detection that is based on various color spaces are an example for this type of detection. In other hand, region based skin detection focus on spatially arrangement of skin pixel with additional information of intensity and texture. However, Vezhnevets et al. [5], Kakumanu et al. [6], and Phung et al. [7] has conducted surveys about skin color modeling and skin segmentation based on color information in different color spaces. Phung has discovered that skin detection accuracy does not depend on choice of color space or color quantization bin sizes. Besides for skin segmentation a few researchers have also used texture [8, 9, 10] or shape [11] information in combination with color-based methods. Wavelets, contourlets [12, 13], and textons are among the textural features that have been used in conjunction with color cues. Convolutional Neural Network (CNN) received mentionable amount of attractions for image classification and object detection. So, there was some approach to classify and pixel-wise prediction using deep neural network architecture. This CNN method outperforms hand crafted threshold based method. But CNN does not work well defining relationship between pixel and its neighbors. Since skin pixels are located close to each other spatially, CNN lacks the desired performances. For all of this previous work, color cues have provided the dominant source of information. In spite of the emphasis on color-based analysis, a considerable number of applications will benefit from a system that can perform skin detection in the absence of color cues. Moreover, previously mentioned approaches heavily depend on color information that can lead to incorrect detection. Despite of the fact that there are various methods are present, it is tough to pick one which would project good accuracy in various conditions. So, it is worth looking to a new technique for better performance.The purpose of this study is to find a skin detection method which does not relay on any predefined thresholds. A skin detection technique must be robust in spite of varying conditions. Moreover, the proposed technique needs to be tested and experimentally evaluated to establish its reliability. The primary contributions of the work areI. Proposed method take advantage of broad spectrum of LAB color space and separate illumination by considering only A and B components. It ensures the pixels characteristics without being biased of light presenceII. K-means clustering is used for clustering. The main challenge for an unsupervised learning is to automatically define number of clusters. Features in images varied greatly and number of cluster should be defined dynamically. This challenge was solved in this study by considering largest areas of connected objects.III. For region growing, seeding points are crucial. Hence, by clustering further it can be ensuring that seeding points obviously belongs to skin pixels. Moreover, standard deviation of only skin pixels is calculated which group maximum skin pixel and provide correct region of Interest of an image.Chapter 02Literature ReviewIn this section, some related works are discussed that is relevant to the study. The first part is skin detection methods followed by the clustering technique used in this study and a cluster validation index. Various methods have been proposed for skin detection. In term of technical merit, they can be divided into two broad categories- statistical based method and dynamic adaptive method [14, 15]. Statistical method relays on skin features that can be derived from training sample [16]. Researchers first try to get the color features of skin pixels for a skin classifier. This approaches used a threshold for various basic color spaces. Moreover, there are many color space available- Basic color space, perceptual color space, Orthogonal color space and perceptually uniform color spaces.2.1 Color Spaces in Skin DetectionColor space is a 3D space with axes appropriately defined for all possible human perceptions. Among different color spaces, some of them are imaging device depended and not close to human vision. This color spaces are modified for digital applications. However, color is not directly utilized in various skin detection methods but it affect the performance of a skin detection algorithm. Albiol et al. [17] has mathematically showed that, skin detection performance is independent from choice of color space. As mentioned earlier, device depended color spaces are popular among researchers for skin detection.RGB is a color space that is derived from the cathode ray tube display and as the name suggests, it comprise with three color-red, green and blue. Brand, J and Mason [18] , Jones and Regh [19], Caetano and Barone[20], Oliver et al. [21], Kim et al. [22], Schwerdt and Crowley [23], Sebe et al. [24], Storring et al. [25],Wang and Sung [26], Yang and Ahuja [27],Yang et al. [28] , used RGB color space for skin detection. But high correlations between channels is a drawback of this color space with mixing chrominance and luminance portion. However, some approaches take the advantage of normalized RGB. In normalized RGB, all the channel valus are normalized. As the sum of all the channels are known, then the third component can be avoided. This technique reduces space dimensionality. Brightness of source RGB deeply depends on red and green channel and normalization can separate it. So, for matte surface the ambient light can be ignored and normalized RGB remain constant for the changes in surface [29]. These advantages attract many researchers on color space based skin detections [30, 31]. But normalized RGB color space suffers from uneven illuminations. CIE (Commission Internationale de lEclairage) system represent the color space with Y as luminance. CIE-XYZ was developed from a psychological experiment and it is close to human visual system [32].Besides, various device dependent color spaces are used for TV transmission and digital photography. The orthogonal color space family includes well-known color spaces such as YCbCr, YCgCr, YDbDr, YPbPr, YIQ and YUV. In these color spaces, unlike the RGB, the luminance channel is separated from chrominance resulting a very tight pseudo-ellipse skin cluster. Due to the particular features of these color spaces such as separation of luminance-chrominance channels, relatively simple RGB conversion and relatively tight skin cluster, orthogonal color models have been frequently used in skin segmentation [33-36]. Perceptual color spaces such as HSI, HSV, HSB, HSL and TSL are very attractive color spaces in skin detection literature. In these color models, each pixel is presented by Hue (tint or color), Intensity (lightness) and Saturation associated with physiological feeling of human [37].An artist idea of using saturation, tint and tone was taken into account. These concepts introduce HSV color space where hue (H) defines the most dominant color of a concern area, saturation represent colorfulness with respect to brightness. However, intensity, lightness or value/ variance defines luminance. The capability of intuitive of color components and ability to distinguish between luminance and chrominance make it popular for skin detection. HSV color space is used by Garcia and Tziritas [38], McKenna et al. [39], Saxe and Foulds, [40], Sobottka and Pitas [41], Thu et al. [42], Wang and Yuan [43], Zhu et al. [44]. Various characteristics of hue- invariant to highlights at white light sources, matte surfaces to ambient light and surface orientation for light source was discussed at [45]. However, Poynton [46] mentions some shortcomings like- discontinuation of hue and brightness computation complexity with color vision properties. Moreover, logarithmic transformation based HS named as Fleck HS introduced by Fleck and used by Zarit [47] .A similar color space named as TSL-tint, saturation and Lightness.