Skin segmentation system automatically adapted to image illumination conditions

• Main Authors: Yi-Ting Wang, 王意婷
• Other Authors: Hung-Ming Sun
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/94786729661413283965

碩士 === 開南大學 === 資訊管理學系 === 97 === Skin detection is a two-class classification problem; every pixel in an image is finally assigned to the skin class or non-skin class based on a decision scheme. Several classification methods can be used to solve this problem, for instance, explicit skin-color thresholding, Bayes classifiers, neural network classifiers and so on. Previous studies of skin detection with Bayes classifiers employed a fixed skin color model to treat each image regardless of the illumination condition in the image. However, presentation of skin color in images can be influenced by illumination conditions and thus makes the result of skin segmentation degrade. To solve this problem, a self-adapting method is proposed, which can conform to varying skin color caused by different lighting conditions or races. This method collects some representative skin samples from the input image and uses them to tune a trained skin color model. The adapted skin color model can match better the skin distribution in the input image and increases the skin segmentation accuracy. The proposed approach consists of a training stage and a skin detection stage. In the training stage, a skin image data set and a non-skin image data set are scanned independently to calculate a skin color model and a non-skin color model. In the skin detection stage, the trained skin color model is used to select the pixels most likely to be skin in an image. Then, the selected skin pixels are used to build a new skin color model and this new model is combined with the trained skin color model to detect all the skin pixels. The underlying idea of the proposed approach is that the trained skin model is tuned by imposing a local skin model to make the resulting skin model adapt to the current image. After testing over 18000 images, it shows that the new method performs better than the original method.