| 摘要 |
<P>PROBLEM TO BE SOLVED: To provide a method which can accurately extract a skin color area even if an object image includes various skin colors. <P>SOLUTION: For each pixel in the skin color area in a sample image, a histogram is generated concerning a hue H and saturation S. Then, the histogram is approximated with a Gaussian mixed model having a plurality of Gaussian models. Then, a color-difference signal (H, S) for a pixel in the object image is converted into one-dimensional information by using a corresponding value in the Gaussian model. Then, the value of the one-dimensional information is changed. The increment of the number of corresponding pixels in the Gaussian model is calculated at that time. The value of the one-dimensional information when the increment is nearly minimum is set as a threshold. A pixel having one-dimensional information included in a range defined by the threshold is selected from the object image. Consequently, the skin color area can be extracted from the object image. Furthermore, by using structural information on the extracted skin color area, a face area can be also selected from the skin color area. <P>COPYRIGHT: (C)2004,JPO&NCIPI |
