The effect of linearization of range in skin detection

There is a wide range of work detailing the effect of colorspace on the detection of skin tone pixels. However, one important consideration which has been overlooked is the effect of dynamic range compression. Specifically, many very commonly used file formats (JPEG, PNG, GIF, etc.) use dynamic range compression to compensate for the range expansion implicit in many display systems. When building a model for skin tone detection, this range compression can be detrimental to the performance of a skin classifier. In this paper, we demonstrate how linearization of the range can aid in the exposure invariance of building a skin model. Two common colorspaces used in the classification of skin pixels are HSV and normalized RGB. We show why linearization of the range is important before transformation into either of these spaces. Experimental results are shown for each colorspace used.