Face facial features, the literature on the mouth of the relative positioning of the eyes is much less in terms of positioning, positioning method is also very different. The author combines problems encountered in this process in accordance with the positioning, briefly described as follows:
Positioned generally divided into the mouth, the beard region is removed, the mouth region acquiring, positioning the mouth
1. beard area removed
Interference beard, the biggest problem is the positioning of the mouth, especially in some countries in the Middle East, bearded man very much. At present, a lot of literature, and did not propose solutions to this problem. "Three Dimensional Modeling Based on Automatic Face single frontal face photograph" Gang Wang mentioned weighting BR G color contrast with methods to remove the beard portion. Specifically, according to "B component (blue) and G components relatively similar (green) is located in lip color, but the color, the distribution of the G component is significantly greater than the B component; in the lip and the skin, the R component (red) and B component distribution is relatively stable, G component are the main elements causing chromaticity distribution difference. "
Fig1 Fig2 beard after the removal of the original image
Actual location, we only need to divide the lower half of the face image can be detected.
2. mouth area to obtain
(1) feature reinforcing Mouth
I think through the mouth and gray skin color differences, to get the mouth area is quite difficult. Because in reality, by the light and individual differences, not everyone's lip color and skin color are obvious differences. Thus the need to highlight features of the mouth region, increase contrast. Fisher uses a method of transformation, were collected two samples (skin and lips), seeking inter-class distance equivalent to the maximum principle, to find the most color and solid color classification matrix.
Fig3 Fisher enhance Fig4 Fisher enhancement
The image is an original image on the row, the row is the effect of enhancing the mouth.
(2) elimination of the influence of light
Since the light conditions are not fixed, it will acquire the mouth region, and a great deal of influence subsequent accurate positioning of the mouth. Author Gabor method to handle to good effect.
Fig5 Gabor transform
(3) mouth region of the binarized
Try using some of the image binarization method to extract the region of the mouth, and comprising OTSU binarization method based on edge information of the feature, but not stable tested. The final method used is similar to the "eye position" mentioned in the article, the effect is stable and works well.
Fig 6 OTSU binarization results
Fig 7 based on the edge information retention binarization
FIG 7 is better than OTSU algorithm, but it is susceptible to mouth posture or the like, resulting in a large area binarized, this corner point subsequent to screening inconvenience.
Fig8 mouth region based on Gabor binarized
3. Positioning mouth
Try to get a "good" mouth region, positioning accuracy of mouth will bring great convenience. In the region of the mouth, the method and the mouth corner detecting template combination, to pinpoint the position of the mouth.
There are many ready-made corner detection methods, such as Susan and Harris corner detection. The authors chose the latter, but there are a lot of candidates after the detection of indecent point, how accurate screening it? We mouth shape, edge information coming judgment added. Because from the true position of the mouth, colleagues along the edge of the right or left, should get the most edge points.
4. The positioning result
Reproduced in: https: //www.cnblogs.com/ImageVision/archive/2012/03/21/2408506.html