If I understood you correctly, you have imaged a human model wearing clothes that are close to skin color and your job is to be able to extract the "real" skin area(s). If that is correct, then you need to use some side information together with your skin tone color cues, texture is the first thing that comes to my head. Then if the subject's face is visible, you can detect facial features. From the detected face skin area, you construct a model whereby you can distinguish other genuine skin areas (arms, legs, etc).

Hope that helps. If, otherwise, my understanding of your problem is wrong then my apologies.

Try to find a tennis ball of only differing weight in a container of tennis balls. Optically this is impossible. - - - About the same applies for your problem.

Maybe it is feasible to "walk" the extremities from the ends (head, hands, feet) - searching for typical patterns of hem lines etc. Regarding the multitude of clothing styles even this will not be 100 % satisfactory (think of lace, mesh, sheers) - yet improving your detection capabilities.

Given freedom of choice I'd check whether an IR image - or any other narrow band color image - would give significant information about clad/non-clad areas. Also skin and clothing might show different properties regarding absorption in some spectral regions.

Dear Vimala.

This is indeed ongoing research in the skin segmentation domain and two approaches are common:

1. Detect the face (or hand or something else) of each individual and extract skin regions from it to improve the skin model. This topic is called "adaptive skin model" and a good paper to start is from Mittal: "Hand detection using multiple proposals" (in BMVC). A recent paper from Prof. Kawulok "Self-Adaptive Skin Segmentation in Color Images" (in CIARP) does not even need a face detector. The skin model is learned on-line. But the idea remains the same. Adopt the skin model to the current picture.

2. Detect the texture of the skin along with its color. Mr Kitani wrote a cvpr paper about "Pixel-level Hand Detection in Ego-Centric Videos". He clustered the images based on its overall appearance and trained for each cluster of images a separate skin model.

I also wrote a paper about skin segmentation: "Color-based skin segmentation: An Evaluation of the State of the Art" (ICIP). But I evaluated color-based methods only (no texture, no adaptation). These methods are really fast and already provide good results. Adding texture improves the detection a bit but always slows down the detection process a lot. Nevertheless. If you want to try e.g. the Bayesian method, you can try my source code (which is written in c++, using opencv and the boost library).