Title: A Modified Approach of Using Ground Truth Samples for Statistical Skin Detection Methods

Year of Publication: Oct - 2015
Page Numbers: 86-90
Authors: M. Abdullah-Al-Wadud , Abdullah S. Alghamdi
Conference Name: The International Conference on Computer, Electronics, and Biomedical Engineering (ICCEBE2015)
- United Arab Emirates


The performances of the skin detection approaches, which works based on statistical models estimated from the training dataset, depend largely on the labeled ground truth data. However, in the available benchmark datasets, the ground truth data for skin include many non-skin pixels as well as the pixels coming from the ambiguous regions falling on the boundaries of the skin regions. Moreover, the ground truth non-skin data are usually irrelevant to the available neighborhood. For such reasons, skin detection performances of the available methods are not always found to be as expected. This paper present a modified approach of using the ground truth datasets provided with the benchmark datasets for skin detection. The proposed approach takes only the ‘pure’ skin pixels and ‘relevant’ non-kin pixels into consideration while building the statistical models. Experimental results also show that the proposed approach increases the performances of different skin detection methods.