基于视觉显著性与肤色分割的人脸检测-《计算机技术与发展》

文章信息/Info

作者:: 鲍小如1 ; 陈瑞2 ; 曹雪虹2 ; 焦良葆3; 1. 南京邮电大学通信与信息工程学院,江苏南京 210003;
2. 南京工程学院通信工程学院,江苏南京 211167;
3. 南京工程学院康尼机电研究院,江苏南京 211167

Author(s):: BAO Xiao-ru 1 ; CHEN Rui 2 ; CAO Xue-hong 2 ; JIAO Liang-bao 3; 1. School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China；
2. School of Communication Engineering,Nanjing Institute of Technology,Nanjing 211167,China；
3. Kangni Electromechanical Institute,Nanjing Institute of Technology,Nanjing 211167,China)

Keywords:: GBVS; L*a*b* color space; visual salient; face detection; Euclidean distance

摘要:: 针对复杂环境下人脸检测受光照、姿态、类肤色等因素的影响,提出了一种基于视觉显著性与肤色分割的人脸检测算法。首先利用基于图论的视觉显著性算法(graph-based visual saliency,GBVS)提取包含目标人脸的显著图,并对该显著图进行阈值分割和形态学等操作得到初步的二值人脸目标区域。然后截取 LFW 数据库中人脸肤色部分并将色彩空间转换至 L*a*b* 色彩空间,计算出人脸肤色的 L , a , b 分量的均值。接着在初步得到的人脸目标区域中计算每个像素点的 a , b 分量与人脸肤色 a , b 分量的欧氏距离,取该欧氏距离在一定区间内的像素点,计算出它们的质心作为人脸中心点。最后为了减少背景,取初步人脸目标区域边缘到中心点最小距离的 80%为半径,截取图像中的准确人脸区域,实现最终的人脸目标检测。在 LFW 数据库上的实验表明,该算法在检测非端正人脸时好于 AdaBoost 方法,且算法简单、速度快、检测率高。

Abstract:: Because the face detection is affected by light,pose,skin color and other factors in complex environment,we propose a new face detection method considering with visual salient and skin color. First,we use the graph based visual salient (GBVS) to extract the saliency map of the target face region which is conducted threshold segmentation and morphological operations for the preliminary binary face marks. Then,we intercept some face skin color map from LFW database,and convert their color space to L*a*b* color space for computing the mean values of skin color，s L , a and b . Next,we calculate the Euclidean distance of each pixel in the preliminary face region and the average value of skin color,and take out those pixels whose Euclidean distance within a certain range and regard their centroid as the face，s center. Finally,we intercept the accurate area of target face with a radius of 80% of the minimum distance from the edge of the preliminary face region to the center to detect faces. The experiments on LFW database show that the proposed algorithm in detecting non upright faces is better than the AdaBoost,and it is simple and fast,with great detection effect.