基于图像差分与卷积深度置信网络的表情识别

Expression recognition based on image difference and the convolutional deep belief network

针对传统的人脸表情识别方法中提取表情特征时没有去除个体性差异及突出表情关键部位的高层次特征,本文提出一种将图像差分与改进的卷积深度置信网络(CDBN)相结合的表情识别方法。首先对人脸表情图片进行裁剪、降维等预处理,之后将各类表情图像与中性表情图像做差分运算提取各类表情的差分图像,为了提取表情关键部位的深层次特征,本文将卷积受限玻尔兹曼机(CRBM)的可见层单元划分为多个区域,分区进行特征学习,并将此CRBM堆叠起来,形成分区卷积深度置信网络(PCDBN),之后将各表情的差分图像作为PCDBN可视层的输入,并利用对比散度算法逐层训练网络,最后在顶层添加softmax分类器作为输出层以实现表情识别。在JAFFE和CK+表情库上的实验结果均达到了95%以上的识别率,扩大训练样本后,在CK+表情库上的识别率可达99%以上。

A new expression recognition method is presented by combining image difference and improved convolutional deep belief network is proposed. Firstly, some pre-processing steps of images are used, such as cropping and intensity normalization. Then facial expressional details are obtained by calculating the difference between the images of basic expressions and neutral expression,which reflect the information irrelevant to identities. To extract the high-level expression features of the key regions in the expression image, the input visible layer of the convolutional restricted boltzman machine （CRBM） is divided into multiple regions, and the features are learned by partitioning. The improved CRBMs are stacked to form a partitioned convolutional deep belief network （PCDBN）. Then the obtained difference expression ima- ges are used as the input of the partitioned convolutional deep belief network,the network is trained by improved contrastive divergence algorithm layer by layer, and the softmax network is added to the top layer as the output layer to form the classification surface. Finally,the partitioned convolutional deep be- lief network which is well-trained is used to identify the expression images from test samples. The exper- imental results on JAFFE and CK＋ expression database can achieve the recognition rate above 95 % all. After expanding the training samples,the recognition rate on CK＋ expression database can reach more than 99%.