Adaptive Skin Color Detection

Skin color is an important feature of human, and it has been widely used in the area such as face tracking, face detection, hand recognition, objectionable images filtering system and so on. In the last decades, skin detection has been widely studied and many improvements have been achieved, however the state-of-the-art of skin detection are far from perfection. 1) Most of proposed methods employ skin models with fixed threshold, therefore are lack of adaption to the various illumination and image contents. 2) All the skin models are fixed offline, therefore lack of pertinence to the individual image. 3) The bothersome confusing backgrounds occurred frequently in images are not payed enough attentions, which play a major role in the false detection rate in the skin detecton.For aforementioned reasons we try to discuss from three aspects to improve the performance of skin detection, including adaptive segmentation thresholds, dynamical skin modeling and confusing backgrounds eliminateion. The main contributions of the thesis includes:1. A large skin images database are set up and conventional skin detection approaches are empirically studied on this database.2. A skin detection method based on adaptive thresholds is proposed, which can find the optimal threshold candidates through combining skin probability distribution histograms (SPDH) and four heuristic clues, then an ANN classifer is used to find the optimal threshold. In order to eliminating confusing backgrounds an image relation operation is also introduced.3. An on-line skin color model is presented. Since a human face has the similar skin color distribution with the human body, we can use the skin pixels in a detected face to build the skin color model of human body dynamically by GMM.4. An effective confusing backgrounds elimination method is proposed. Through applying skin probability variance histogram and GMM, we can eliminate confusing backgrounds in images.