基于深度学习的微表情特征提取算法设计与实现

发布时间：2018-05-19 17:46

本文选题：微表情识别 + 特征提取　；参考：《北京交通大学》2017年硕士论文

【摘要】：国内外的个人极端行为、危及公共安全的事件呈上升态势,如网络谣言、公交纵火和驾车冲撞敏感区域等。为了对危险行为预警,相关组织和人员开始研究自动预警技术。表情是人类表达情感的重要非言语行为,可作为危险行为预警过程的重要线索。目前针对表情的研究虽已取得了一些成果,但关注的多是普通表情。除了普通表情,还存在难以被觉察的微表情,其持续时间非常短,与潜在意图关系密切,这种表情即为微表情。针对微表情特征提取是一项交叉性的研究课题,涉及计算机、信号与信息处理和临床心理学等多个学科,具有重要的理论研究和实际应用意义,有助于促进各研究领域的相互交流和推进相关技术的发展。本文重点研究了基于深度学习的微表情特征提取算法,对微表情的激活度(Arousal,情绪是觉醒还是昏睡的程度)、效价(Valence,情绪表现积极还是消极)、期望度(Expectation,情绪是惊奇的程度)、强度(Power,受外界影响时控制自己情感的程度)四个情感属性进行了预测分类。最后,将预测值经过一个一维中值滤波进行规整。论文的主要工作包括:(1)提出了一种基于卷积神经网络(CNN)的微表情特征提取算法。与传统的特征提取方法(基于梯度的特征提取算法(HOG)、基于局部纹理的特征提取算法(LBP))相比,本文算法所采用的卷积神经网络架构对人脸表情表现比较集中的地方,如眼角、嘴角等部位激活了更多节点,这样一方面能够从原始数据中学习到具有较高表现力的微表情描述特征,另外可使算法性能不依赖于精确的面部检测和定位过程。(2)提出了一种基于深度学习的微表情情感因素预测及分类算法。所提算法使用多层感知机(MLP)代替CNN中的全连接层,即将CNN所提取到的全局特征连接到MLP进行训练和识别,从而对微表情的激活度、效价、期望度和强度四个情感属性的预测和分类。在AVEC2012微表情库上的实验结果表明,激活度、效价、期望度和强度这四个属性的Topl平均识别率分别为71.51%、73.14%、66.43%、69.05%。
[Abstract]:Personal extreme behavior at home and abroad, public safety incidents are on the rise, such as Internet rumors, public transport arson and driving in sensitive areas and so on. In order to warn of dangerous behavior, relevant organizations and personnel began to study automatic warning technology. Expression is an important non-verbal act to express human emotion and can be used as an important clue in the process of warning dangerous behavior. Although some achievements have been made in the study of facial expressions, most of them focus on ordinary expressions. In addition to ordinary expressions, there is also an imperceptible microexpression, which lasts for a very short time and is closely related to the underlying intention. Microfacial expression feature extraction is a cross research subject involving computer, signal and information processing, clinical psychology and so on. It has important theoretical research and practical application significance. It helps to promote the exchange of research fields and promote the development of related technologies. This paper focuses on the algorithm of micro-expression feature extraction based on depth learning. Arousal.Arousal, the degree to which emotion is awakened or drowsy, the value of value, the positive or negative emotional performance, the degree of expectation, the degree of surprise, the intensity of power, the degree of control of one's emotions when influenced by the outside) Affective attributes are classified as predictors. Finally, the predicted value is regularized by a one-dimensional median filter. The main work of this paper includes: 1) A novel algorithm for feature extraction based on convolution neural network (CNN) is proposed. Compared with the traditional feature extraction methods (gradient based feature extraction algorithm and local texture based feature extraction algorithm), the convolutional neural network architecture used in this algorithm is more concentrated on facial expression, such as the corner of the eye. The corners of the mouth and other parts activate more nodes, so on the one hand, can learn from the raw data with a more expressive micro-expression description features, In addition, the performance of the algorithm does not depend on the accurate facial detection and localization process. (2) A prediction and classification algorithm for micro-expression emotion factors based on in-depth learning is proposed. The proposed algorithm uses multilayer perception machine (MLP) instead of the full connection layer in CNN, that is, the global features extracted by CNN are connected to MLP for training and recognition, so that the activation degree and titer of micro-expression can be obtained. The prediction and classification of the four emotional attributes of expectation and intensity. The experimental results on AVEC2012 microfacial expression database show that the average recognition rate of activation, titer, expectation and intensity of the four attributes is 71.51 and 73.14, and 66.43 and 69.05, respectively.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】