基于多分类模型加权投票法的人脸微笑检测-《计算机技术与发展》

文章信息/Info

Title:: Facial Image Smile Detection Based on Multi－class Model Weighted Voting

Author(s):: FENG Ze－an; WANG Peng; School of Electronic Information Engineering，Xi’an Technological University，Xi’an 710021，China

Keywords:: smile detection; face image; texture; geometry; classification model; weighted voting

摘要:: 为了进一步提高人脸微笑检测率并解决微笑检测系统用于训练标签数据不足的问题，结合人脸图像的纹理和几何特点，应用了一种基于多分类模型加权投票法的微笑检测方法。在经预处理和直方图均衡化的面部图像上，利用局部二进制模式(LBP)和 Gabor 小波变换提取局部的、抽象的特征，同时以人脸特征点检测作为补充，构建了四种不同的分类模型(UPN、GPS、AdaBoost 和 LDA)，分别对人脸图像进行分类检测，同时结合各模型之间互补和各自对微笑检测的优势，通过计算权值对各结果进行加权投票，得到面部图像的最终检测结果。实验结果显示出该方法的有效性，在公开的 GEN-KI－4K 人脸数据集上获得了 95．8%的微笑检测率，比单个分类模型的平均检测率提高了 10．3%，与该数据集的最新的微笑检测率相等。

Abstract:: In order to further improve the face smile detection rate and solve the problem that the smile detection system is not enough for training tag data，we apply a smile detection method based on multi－classification model weighted voting in combination with the texture and geometric characteristics of human face image． On the facial image with pre－processing and histogram equalization，local and abstractfeatures are extracted by the local binary model (LBP) and Gabor wavelet transform． At the same time，facial feature point detection isused as a supplement to construct four different classification models (UPN，GPS，AdaBoost and LDA)． Combined with the complementarity between the models and their respective advantages on smile detection，the final detection results of facial images are obtained bycalculating the weights and voting the results weighted． The experiment shows the effectiveness of this method． The detection rate of95. 8% is obtained on the open GENKI－4K face dataset，which is 10．3% higher than the average detection rate of a single classificationmodel and the same as the latest smile detection rate．