【作者】 郭哲； 袁博雅； 徐芳达； 王毅； 樊养余；

【Author】 GUO Zhe;YUAN Boya;XU Fangda;WANG Yi;FAN Yangyu;School of Electronics and Information,Northwestern Polytechnical University;

【机构】 西北工业大学电子信息学院；

【摘要】 人脸表情分析技术是人机交互、智能控制等领域的热门课题。针对传统人脸表情分析方法需要人工定义提取特征的问题,本文提出了基于深度学习的人脸表情分析方法,分别以VGG19、ResNet18和Dense Net三种不同的卷积神经网络对人脸面部表情特征进行训练学习,并对网络进行微调,有效提高了算法模型的准确率和适应性。在FER2013和The Extended Cohn-Kanade Dataset两个国际公开数据库上的实验结果表明,基于深度学习的表情分析方法对高兴、惊讶、中性三种表情识别率均较高,最高达到了99%;而对恐惧、悲伤、厌恶等相对容易混淆的表情识别率偏低,最低不到50%。更多还原

【Abstract】 Face expression analysis is a hot topic in the fields of man-machine interaction and intelligent control.Aiming at the problem that the traditional face expression analysis method needs to manually define the feature extraction,a facial expression analysis method based on depth learning is proposed in this paper.In this method,the facial expression features are trained and studied with three different convolution neural networks of VGG19,ResNet18 and Dense Net,respectively.And the accuracy and the adaptability of the algorithm model are effectively improved by fine tuning.The experimental results on the two international public databases of FER2013 and The Extended Cohn-Kanade Dataset show that the proposed facial expression analysis method based on depth learning is very good,and the recognition rate of three expressions(happy,surprise and neutral) is high,and the highest recognition rate is 99%;and for the relatively easy-to-confuse expressions,such as fear,sadness and disgust,the recognition rate is low,and the lowest is less than 50%.更多还原