计算机工程

• •    

基于多尺度分层双线性池化网络的细粒度表情识别

  

  • 发布日期:2021-01-05

Fine-grained expression recognition based on multi-scale hierarchical bilinear pooling network

  • Published:2021-01-05

摘要: 由于人脸表情细微的类间差异和显著的类内变化使得人脸表情识别困难而导致识别率低,为此提出 了一个基于多尺度双线性池化神经网络的模型。首先,通过精心设计的三种不同的粗细尺度网络提取人脸 表情全局特征。然后,引入了分层双线性池化层,集成多个同一网络以及不同网络的多尺度跨层双线性特 征来捕获不同层级间的部分特征关系,从而增强模型对面部表情微妙特征表征的判别能力。最后,通过逐 层反卷积融合多层特征信息,解决了神经网络通过多层卷积层、池化层提取特征时丢失部分关键特征的问 题。该模型在 FER2013 和 CK+公开数据集上识别率最高达到 73.725%、98.28%,实验结果表明,所提出的 方法能有效提高表情识别率,优于 SPLM、CL、JNS 等较新人脸表情识别算法。

Abstract: Because the subtle differences between classes and significant changes within classes of facial expressions make it difficult to recognize facial expressions, resulting in low recognition rate, a model based on multi-scale bilinear pool neural network is proposed. Firstly, the global features of facial expressions are extracted through three carefully designed networks with different scales. Then, a hierarchical bilinear pooling layer is introduced, and multiple multi-scale cross-layer bilinear features of the same network and different networks are integrated to capture some feature relationships between different levels, thus enhancing the discrimination ability of the model on subtle features of facial expressions. Finally, by deconvolution and fusion of multi-layer feature information layer by layer, the problem of losing some key features when neural network extracts features through multi-layer convolution layer and pool layer is solved. The recognition rates of this model on FER2013 and CK+ public data sets are up to 73.725% and 98.82%. The experimental results show that the proposed method can effectively improve the facial expression recognition rate, and is superior to SPLM, CL, JNS and other new facial expression recognition algorithms.