Skin-representative region in a face for finding real skin color

Skin detection is used in applications in computer vision, including image correction, image–content filtering, image processing, and skin classification. In this study, we propose an accurate and effective method for detecting the most representative skin color in one’s face based on the face’s center region, which is free from nonskin-colored features, such as eyebrows, hair, and makeup. The face’s center region is defined as the region horizontally between the eyes and vertically from the middle to the tip of one’s nose. The performance of the developed algorithm was verified with a data set that includes more than 300 facial images taken under various illuminant conditions. Compared to previous works, the proposed algorithm resulted in a more accurate skin color detection with reduced computational load. Introduction With the development of information technology and mobile devices, various kinds of vision-based applications have been developed. Among those applications, many facilitate skin detection algorithms. A skin detection algorithm is a process of detecting skincolored pixels and regions in a digital image. Skin detection is considered a key step because the color of human skin is cognitively highly relevant and, accordingly, it is an effective feature for computation. The main applications of skin detection are image correction, image–content filtering, image processing, and skin classification [1]. For example, hand gestures can be classified after detection and segmentation of the skin region [2–4]. A human tracking algorithm and pornographic image filtering also uses skin detection before tracking and filtering process [5–7]. Not limited to these purposes, a number of skin detection algorithms have been developed. The approach of skin detection is largely classified by two types. The first type is a pixel-based skin detection method. Skin color mainly differs in lightness, but it does not show significant differences in hue and saturation, even across different ethnic groups. Likewise, skin colors are distributed in a narrow range of color spaces [8], so it is possible to screen pixels that should not be included in the skin color range. Because skin pixels should belong to the predefined range, the computational load could be relatively low. Many previous studies followed the skin range assumption, and the range is defined differently depending on the color spaces, such as RGB, normalized RGB, HSV, YCbCr, YUV, and CIE L*a*b* [9, 10]. However, the pixel-based skin detectors did not work properly for the color shifts caused by various illuminant conditions. In particular, there were high color shifts for facial images taken under various correlated color temperatures or under low illuminance. In addition, these methods cannot distinguish nonskin regions, which have similar colors to skin. As presented in Figure 1, brown hairs on the subject’s forehead were detected as skin due to the similar color range. Furthermore, the images captured by camera can be distorted due to the illuminant and the characteristics of each device. As mentioned, the range of skin color is relatively narrow and the distortion may lead the algorithm not detecting the skin region correctly [11]. This type of existing pixel-based skin detector may filter out the skin region if it is distorted due to extreme chroma of the illuminant. One possible solution is to correct images first, which is known as color constancy. For example, some assumptions were suggested, such as the average color of the scene being gray [12] or the brightest pixel in the scene being white [13]. However, these assumptions are not truly operational in some cases, and we are not always able to control the illuminant. Figure 1. The pixel-based skin detection often fails because it does not discriminate nonskin regions as far as hues belonging to the predefined detection boundary. In this case, the brown hairs were detected as skin. The other approach of skin detection is an adaptive skin detection algorithm. This does not use a predefined detection boundary of skin regions. This takes the spatial arrangement of pixels into account; therefore, it has more flexibility in various illuminant [7, 12, 13]. Most of these methods detect the face first to extract the skin region. After facial recognition, the algorithm detects skin regions by calculating the dominant hue in the face, so it can detect the skincolored pixels near the dominant color in the color space [14]. In this way, the adaptive skin color detection technique computes skin pixels in a device-dependent color space that works properly under various illuminant conditions. However, this method inevitably includes features that are not skin colored, such as eyebrows, hair, and makeup. These nonskin-colored regions may cause inaccurate results for further process after skin color detection. For example, an estimation of accurate skin color is rarely possible if the nonskincolored region is included. In addition, this type of algorithm has more computational load than the pixel-based skin detection. As skin detection is a primary process for many of other applications, the computational load can be an important issue. As such, an improvement in skin detection is necessary to meet users’ needs for the development of vision-based. In this regards, it is anticipated to obtain the most representative skin color in one’s face to detect the other skin regions in the image without considering the nonskin features of the face. In addition, it will be more effective if the representative skin color could be calculated instead of taking the whole face region for computation. 66 IS&T International Symposium on Electronic Imaging 2017 Computational Imaging XV https://doi.org/10.2352/ISSN.2470-1173.2017.17.COIMG-425 © 2017, Society for Imaging Science and Technology