基于深度卷积神经网络的脸部表情分类研究

Research on facial expression classification based on deep convolutional neural network

为了更加精确地对人脸表情进行分类,文中提出了使用CNN(卷积神经网络)来进行脸部表情的识别分类,其包含8层网络,前5个是卷积层(C1-5),其余3个是全连接层(FC6-8)。最后一个全连接层的输出提供给6路softmax,其在6类标签上产生一个分布。文中收集各种数据库,并将数据库组织成6个表情类,如"中性"、"高兴"、"伤心"、"愤怒"、"惊讶"和"反感"。应用预处理和数据增强技术来提高培训效率和分类性能。调整卷积层的特征映射的数量与全连接层的节点数,找到最能表达6个面部表情特征的最优结构。并通过交叉验证和交叉数据库实验得出,文中提出的CNN结构具有良好的脸部表情分类性能。其次,与其他传统模型相比,所提出的CNN结构在分类性能方面具有优越性且执行时间更短。

In order to classify facial expressions more accurately,this paper proposes to use CNN(Convolutional Neural Network)to identify and classify facial expressions,which consists of eight layers of networks,the first five are convolutional layers(C1-5).The remaining three are fully connected layers(FC6-8).The output of the last fully connected layer is provided to 6 softmax,which produces a distribution on the 6 types of labels.This article collects various databases and organizes the database into six expression classes,such as"neutral","happy","sad","anger","surprise"and"dislike".Apply pre-processing and data enhancement techniques to improve training efficiency and classification performance.Adjust the number of feature maps of the convolutional layer and the number of nodes in the fully connected layer to find the optimal structure that best expresses six facial expression features.Through cross-validation and cross-database experiments,the proposed CNN structure has good facial expression classification performance.Secondly,compared with other traditional models,the proposed CNN structure is superior in classification.