多人合影图片中欢乐气氛的估计
Estimation of happiness intensity in group picture with multiple people
  
DOI:
中文关键词:  面部表情识别;情感识别;卷积神经网络;人群高兴程度
英文关键词:facial expression recognition;emotion recognition;convolutional neural networks;group happiness intensity
基金项目:国家自然科学基金(61501249)和江苏省重点研发计划(BE2016775)资助项目
作者单位
卢官明 南京邮电大学 通信与信息工程学院,江苏南京210003 
张文静 南京邮电大学 通信与信息工程学院,江苏南京210003 
闫静杰 南京邮电大学 通信与信息工程学院,江苏南京210003 
摘要点击次数: 3091
全文下载次数: 1245
中文摘要:
      为了描述多人合影图片中的欢乐气氛,提出了一种基于卷积神经网络的估计方法。首先,采用多任务级联卷积神经网络(MTCNN)对合影图片中的所有人脸进行检测。然后,采用深度卷积神经网络对检测出的每一个人脸分别进行高兴程度的估计。最后,对被检测到的所有人脸的高兴程度进行加权平均来估计合影图片中的欢乐气氛。在HAPPEI数据库验证数据集上的实验结果表明,提出的方法具有良好的性能,优于使用Census变换直方图(CENsus TRansform hISTogram,CENTRIST)特征描述子和非线性支持向量回归相结合的方法,前者的均方根误差(RMSE)为0.623 4,后者的RMSE为0.776 8。
英文摘要:
      To describe the happiness intensity in group picture with multiple people, an estimation method based on convolutional neural network is proposed. Firstly, the multi task cascaded convolutional neural networks (MTCNNs) are used to detect all faces in a group picture. Then, the face level happiness intensity of each detected individual face is respectively estimated by using a deep convolutional neural network. Finally, the weighted average of face level happiness intensities for all detected faces in the group picture is used to estimate the happiness intensity of a group of people. Experimental results on the validation dataset of HAPPEI database demonstrate that the method has good performance with root means quare error (RMSE) of 0.623 4 , and outperforms the baseline approach by using the feature descriptor of CENsus TRansform hISTogram (CENTRIST) and support vector regression with a non linear kernel, having an RMSE of 0.776 8.
查看全文  查看/发表评论  下载PDF阅读器

你是第3755956访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计

欢迎访问《南京邮电大学学报(自然科学版)》编辑部!