基于深度学习的表情识别方法研究

发布时间：2019-03-11 20:40

【摘要】：人脸表情是人类交流的一种必不可少的方式,通过对人脸表情的研究可以探索人类的心理状态,进而充分的了解人们的行为意图。深度学习是一种特征学习方法,通过一些简单的非线性的模型把数据转变成为更高层次的,更加抽象的表达,可以解决语音处理、计算机视觉、自然语言处理等问题。本文利用深度学习来解决表情识别中的一些问题,并通过实验进行了验证。本文的主要研究内容如下:1.本文研究了多种类型的深度学习模型,按结构可分为深度卷积神经网络、深度信念网络、深度玻尔兹曼机、堆叠自动编码器和递归神经网络等。它们算法各异,且所适用领域也不相同。因此,选择合适的深度学习模型是解决表情识别问题的关键。本文经过对比论证发现,深度卷积神经网络其独特局部连接与权值共享机制可以解决表情识别中特征维数大,计算困难等问题。因此,本文选择深度卷积神经网络作为本文的深度学习模型。2.针对静态表情识别中提取特征会损失图像原有的特征信息的问题,提出利用深度学习模型中的深度卷积神经网络实现表情特征提取。由于深度卷积神经网络避免了对图像的复杂前期预处理,可以直接输入原始图像,它通过卷积与池化的共同作用进行特征提取,不需要人为提取特征,并且网络易于训练,比那种全连接的神经网络的泛化性能更好,因此将深度卷积神经网络应用静态表情识别中。3.针对动态表情识别中系统抗干扰性差,运算速度慢以及实时性差等问题,提出利用深度卷积神经网络实现动态表情特征提取。由于动态表情识别系统输入的实时获取的动态人脸表情序列,与静态表情识别不同,这需要系统实时将获取到的人脸进行存储和识别。为了解决这一问题,利用Haar分类器进行人脸检测,再引入深度卷积神经网络,通过它构建图像的本质特征,提取表情特征,最后运用Softmax分类器实现表情分类。4.为了提高深度卷积神经网络的非线性表达能力,从而实现更好地表情特征提取,对网络结构进行改进,利用深度连续卷积神经网络来实现表情识别。本文对深度卷积神经网络进行改进,引用多层小尺度卷积代替单层大尺度卷积的思想,即利用2层连续卷积层来代替单层卷积层,提高网络的非线性表达能力,再对网络的激活函数及参数优化方法做出调整,来提高网络的表情特征拟合能力。
[Abstract]:Facial expression is an indispensable way of human communication. Through the study of facial expression, we can explore the psychological state of human beings, and then fully understand the behavior intention of human beings. Deep learning is a feature learning method. It can solve the problems of speech processing, computer vision, natural language processing and so on by transforming data into higher level and more abstract expression through some simple nonlinear models. In this paper, deep learning is used to solve some problems in expression recognition, and it is verified by experiments. The main contents of this paper are as follows: 1. In this paper, many kinds of deep learning models are studied, which can be divided into deep convolution neural network, depth belief network, depth Boltzmann machine, stacking automatic encoder and recurrent neural network. They have different algorithms and different fields of application. Therefore, choosing the appropriate deep learning model is the key to solve the problem of expression recognition. Through comparison and demonstration, it is found that the special local connection and weight sharing mechanism of deep convolution neural network can solve the problems of large feature dimension and difficult computation in facial expression recognition. Therefore, in this paper, the deep convolution neural network is chosen as the depth learning model of this paper. 2. In order to solve the problem that feature extraction in static expression recognition will lose the original feature information of image, this paper proposes to use depth convolution neural network in depth learning model to realize expression feature extraction. Because the deep convolution neural network avoids the complex pre-processing of the image, it can directly input the original image. It can extract the features through the joint action of convolution and pooling, and it does not need man-made feature extraction, and the network is easy to train. The generalization performance of the fully connected neural network is better than that of the fully connected neural network, so the deep convolution neural network is applied to static expression recognition. In order to solve the problems of poor anti-jamming, slow computing speed and poor real-time performance in dynamic expression recognition, a method of dynamic expression feature extraction based on deep convolution neural network is proposed in this paper. Because the real-time acquired dynamic facial expression sequence from the dynamic expression recognition system is different from the static facial expression recognition, it is necessary for the system to store and recognize the acquired face in real-time. In order to solve this problem, the Haar classifier is used for face detection, and then the deep convolution neural network is introduced to construct the essential features of the image, extract the expression features, and finally use the Softmax classifier to realize the expression classification. 4. In order to improve the nonlinear expression ability of the deep convolution neural network and achieve better expression feature extraction, the network structure is improved and the deep continuous convolution neural network is used to realize the expression recognition. In this paper, the deep convolution neural network is improved, and the idea that the multi-layer small-scale convolution replaces the single-layer large-scale convolution is introduced, that is, the two-layer continuous convolution layer is used to replace the single-layer convolution layer, and the nonlinear expression ability of the network is improved. Then the activation function and parameter optimization method of the network are adjusted to improve the expression feature fitting ability of the network.
【学位授予单位】：长春工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】