基于深度学习的人脸表情识别

发布时间：2018-01-06 02:17

本文关键词：基于深度学习的人脸表情识别　出处：《浙江理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：人脸表情识别是当前计算机视觉、模式识别、人工智能等领域的热点研究课题。它是智能人机交互技术中的一个重要组成部分，近年来得到广泛的关注，不同领域的研究者提出了许多新方法。本文综述了国内外近年来人脸表情识别技术的最新发展状况，对人脸表情识别系统所涉及到的关键技术：人脸表情特征提取和人脸表情分类，分别做了详细的分析和归纳。最后，总结了人脸表情识别的研究现状，并指出了其未来的发展方向。本文主要研究了在人脸表情识别中特征提取和分类中的一些关键问题，并结合深度学习的方法提出了一些改进方法，最后通过实验进行了验证。本文的主要工作如下: 1.提出一种融合深度信念网络和多层感知器的人脸表情识别新方法。首先采用深度信念网络对提取的原始人脸表情图像的初级特征或局部二元模式（LBP）征进行无监督学习，得到更高层次的抽象特征，然后将其用于初始化多层感知器模型中的隐层网络权重值，最后利用该初始化后的多层感知器实现人脸表情的分类。在JAFFE数据库中，该方法能够取得最好的人脸表情正确识别率为90.95%；在Cohn-Kanade数据库中，，取得了最好98.57%的人脸表情正确识别率。而且与其它识别方法相比，深度信念网络（DBNs）方法有着明显的优势。可见，该方法用于人脸表情识别，可以较好地改善识别性能。 2.对深度信念网络的鲁棒性人脸表情识别性能做了研究。考虑到在人脸表情识别过程中图像可能受到噪声的影响，在对测试图像存在像素腐蚀的情况下，着重对基于深度信念网络的鲁棒性人脸表情识别性能进行了探讨。深度信念网络具有很强的无监督学习的能力，在不同的腐蚀比例下，仍然能取得不错的识别效果。在Cohn-Kanade数据库中，实验结果表明DBNs具有优越的分类性能和鲁棒性，是非常适合于人脸表情识别的。 3.设计了人脸表情识别的GUI界面。在完成人脸表情识别的程序设计后，根据GUI系统设计的简单性、一致性、习常性，设计了人脸表情识别的GUI界面，方便程序的操作使用。
[Abstract]:Facial expression recognition is the current computer vision, pattern recognition, hot research topics in the field of artificial intelligence. It is an important part of intelligent human-computer interaction technology, has received wide attention in recent years, researchers in different fields and put forward many new methods. This paper reviews the technology of facial expression recognition in recent years. The latest development status and key technologies involved in the system of facial expression recognition, facial feature extraction and facial expression classification, were analyzed and summarized in detail. Finally, summarizes the research status of face recognition, and points out the future direction of development.
This paper mainly studies some key problems in feature extraction and classification of facial expression recognition, and proposes some improvement methods combined with deep learning method. Finally, it is verified by experiments.
1. this paper proposes a new method of facial expression recognition fusion depth of belief network and multilayer perceptron. Firstly, the primary characteristics of the two modes of local or deep belief networks to extract the original facial expression image (LBP) features for unsupervised learning, get a higher level of abstraction features, and then applied to the hidden layer weights initialize the multi-layer perceptron model in the value of the final realization of facial expression classification using multilayer perceptron. After the initialization in the JAFFE database, this method can achieve the best correct facial expression recognition rate was 90.95%; in the Cohn-Kanade database, obtained the correct recognition rate of 98.57%. The best facial expression and compared with other methods. Deep belief network (DBNs) method has obvious advantages. Obviously, the method for facial expression recognition, can effectively improve the recognition performance.
Robust facial expression recognition performance of 2. deep belief networks have been studied. Considering the influence of image by noise in facial expression recognition process, in the presence of corrosion on the pixel test image case, focuses on the robustness of face recognition based on the performance of deep belief networks are discussed. The deep belief network has no the ability of supervised learning is very strong, in corrosion under different proportion, still can achieve good recognition effect. In the Cohn-Kanade database, the experimental results show that DBNs has excellent classification performance and robustness, is very suitable for facial expression recognition.
3., the GUI interface of facial expression recognition is designed. After finishing the program design of facial expression recognition, according to the simplicity, consistency and habit of GUI system design, the GUI interface of facial expression recognition is designed to facilitate the operation and application of programs.

【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】