基于局部梯度和面积重叠合并法的人脸检测

Face Detection Based on Local Gradient Patterns and Square Overlap Merge Method

研究了一种人脸检测方法。采用局部梯度模式(Local Gradient Patterns,LGP)提取人脸特征,用AdaBoost学习算法进行了层级分类器的训练。提出了应用面积重叠合并的识别方法(Square Overlap Merge Method,SOMM),可以降低检测错误的正检测误差,同时克服了训练样本少的情况下分类器可靠性差的缺点。实验中采用MIT的人脸数据库进行分类器的训练,并在训练好的分类器的基础上,又进一步采集了6000多张人脸图片,进行分类器的再训练,以求分类器准确可靠。实验证明上述方法能够快速有效的检测人脸,并且能很好的克服光照、姿态、背景、遮挡物等对人脸检测的影响。

The problem of face detection was discussed in this paper. The features of faces were extracted by using Local Gradient Patterns（LGP）. AdaBoost way was used to train cascaded classifiers. Square Overlap Merge Method （SOMM） was exploited to detect face in order to reduce the false positive detection error greatly, which can make up the disadvantages of the low reliability of classifiers trained by a small amount of training samples. In the ex- periment, the MIT databases were utilized to train the classifiers. In order to make sure the classifiers more accurate and reliable, we collected additional 6000 face images to train the classifiers trained before. The results of the experiment show that this method can detect faces fast and effectively. Furthermore, it can overcome the impact of illumination, posture, background, occluder and etc.