Robust Skin ​Detector based on face pixels.

​

The skin detection results are enhanced by using detected face regions (using the OpenCV face detector) to determine the color of skin in the target image. This enables skin to be detected despite unusual lighting and color balances which may be peculiar to the target image. A simple classifier of color likelihood is used based on a histogram of the face pixels. We proceed in two stages. First, in the manner of hysteresis in the Canny operator, two thresholds (confidences) for likelihood are learnt. Pixels which are above the high threshold are classified as skin. Then pixels which are above the low threshold are also classified as skin if they are spatial neighbors of a pixel above the high threshold (these thresholds are learnt by cross-validation on ground truth segmentations). Second, a bootstrapping stage, the colour of the neighbouring pixels is used to update the color likelihood classifier and the process is repeated.