Accurate identification of skin regions in images is a vital step in many applications of image processing. Face detection, image understanding, blocking of offensive content--all utilize skin classification. As a result, there is a significant body of work that explores a number of techniques. The authors provide a summary of the most important findings to date and comparatively test the leading methods.

Khan et al. define four elements for identifying the ideal components of a skin detection application: color space, illumination compensation, skin color modeling, and color constancy algorithms. Each element is addressed by a significant body of work that is well covered and well referenced. The methodology here is to apply methods in each element area to a manually labeled set of images containing skin regions. Elements are evaluated in combination; that is, each color space is tested with each proposed skin color model. Results are evaluated by comparison of the F-measure statistic.

Previous work on color spaces tends to favor cylindrical spaces such as hue, saturation, and value (HSV). The authors find that the best overall space for accurate skin detection is improved hue, luminance, and saturation (IHLS), which has been identified as advantageous for other color segmentation applications. IHLS has the best performance for nearly all color modeling methods, while normalized red, green, and blue (RGB) has the worst performance for all color models. Illumination compensation attempts to eliminate variations caused solely by inconsistent scene illumination. Some researchers remove all luminance information, which the authors have confirmed to be overly drastic. Their results show that retention of luminance is beneficial when using a good color space such as IHLS or hue, saturation, and intensity (HSI).

“Skin color modeling” is their term for the classification method used to operate on the pixel-level data. Each color pixel is segmented individually, without use of information from neighboring pixels. The methods evaluated include many leading classification techniques: multilayer perceptron (MLP), random forest, and Bayesian classification and support vector machines. While results vary by color space, the best three techniques are random forest, MLP, and a decision tree method known as J48. While differences in F-measure were significant (random forest performance was nearly three times that of the AdaBoost classifier), it is likely that the different algorithms had different degrees of optimization--a careful application of a support vector machine (SVM) may well outperform a rudimentary application of random forest. It should also be noted that there are other ways to compare results than by using the simple F-measure of classification accuracy. Some techniques tended toward false positives (including nonskin areas), while others might reject actual areas of skin.

Finally, the use of color constancy methods is explored. The notion of these algorithms is that some in-scene properties can be extracted to estimate the chromatic nature of the scene illumination. For example, the gray-edge method considers that differences in reflectance in a scene will not affect the color of the light reflected, only its intensity. The effect of such known illumination variations is to “stretch” the skin region in the color space in a known manner; this gives rise to a computable correction that can be applied to the original scene. Overall, the authors’ results confirm that this correction can increase performance, though they only generated results for the YCrCb space and the random forest classifier.

In conclusion, this paper is a concise summary of methods for achieving skin detection in color images. The authors compare leading approaches to the elements of color space selection, color modeling, luminance, and illumination correction. The paper is reasonably well written and quite approachable. All of this work builds on color understanding and processing in the larger realm of general image processing, and more consideration of this larger body of work would be helpful. Still, this paper serves as a useful comparison of a wide range of methods; the authors’ work to compare different elements in all combinations is especially helpful.