Automatic Image Grading Based on Skin Segmentation

This paper proposes an automatic image grading method, which classifies an image into three levels, i.e., Normal, Revealing Attire and Nude. First, a novel region based skin detection method, which incorporates the clues of color, shape, texture and neighborhood, is used to get the skin regions. Then a normalized mask is generated from the skin region image according to the scale and location of the face. Global and spatial features extracted based on this mask are used as the input of SVM to give the grade of an image. Besides, because false classifications of images with different grades have quite different affections, a cost-matrix is defined and the MetaCost method is used to get the minimum-risk results. Experimental results show the effectiveness of our method.