基于重建图像信噪比特征的脸部位置检测方法

Detection of Face Position Based on SNR of Reconstructed Images

目的通过研究找到一种基于重建图像信噪比(signal-to-noise,SNR)的人脸检测方法,从而提高在图片中找到人脸所在位置的准确率。方法首先通过图像向特征脸空间投影得到重建图像,然后利用重建图像的SNR进行人脸检测。经实验发现,在对一幅图像进行扫描的过程中,人脸的位置既是信噪比值横向的极大值点,又是纵向的极大值点,且在单幅人脸图像中,人脸处的SNR为全局极大值,因此可以利用该动态规律准确地找到人脸位置。结果利用上述方法对耶鲁人脸库100张人脸和自拍的50张人脸进行实验,结果表明,通过搜索全局最大值确定出人脸的位置,准确率为98%。进一步,利用上述方法对已经得到的人脸进行第二次搜索,找到不包含头发等周围图像的中心脸部区域。最后,通过图像锐化和模板匹配相结合的方法找到眼睛位置,旋转图像使双眼在同一水平位置上,并根据比例关系可重新精确地划出中心人脸区域,眼睛定位准确率达96%。结论基于重建图像SNR的人脸检测方法可以提高寻找人脸的准确率,因此该方法是一种简单而有效的脸部位置检测方法。

Objective A new detection method of face position based on signal-to-noise （SNR） of reconstructed images was developed, which can improve the accuracy of finding the face position in the image. Methods The SNR of reconstructed images was acquired by projecting to eigenface space and used in the face detection. Correspondingly,the face was detected according to the dynamic change of SNR. The results of experiments showed that the SNR of faces in whole image was the maximum when the image was scanned horizontally and vertically. If the image only includes one face, then SNR of the face was global maximum. Results One hundred images from face database of Yale University and 50 images from photos acquired by camera were detected. The correct rate of the detection reached to 98%. Furthermore,we scaned the acquired faces by above method again, and then the center zone of face was marked without hair and so on. In this face, the positions of eyes were determined by sharpening and template matching. The face would be rotated in order to make eyes being horizontal, then the face were cropped again according to the proportions. The correction rata of eye position detection reached to 96%. Conclusions The detection method based on SNR of reconstructed images improves the accuracy of finding the face position in the image, and it is simple and efficient for face position detection.