摘要： 人脸表情细微的类间差异和显著的类内变化增加了人脸表情识别难度。构建一个基于多尺度双线性池化神经网络的识别模型。设计3种不同尺度网络提取人脸表情全局特征，并引入分层双线性池化层，集成多个同一网络及不同网络的多尺度跨层双线性特征以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细微特征的表征及判别能力。同时，使用逐层反卷积融合多层特征信息，解决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。实验结果表明，该模型在FER2013和CK+公开数据集上的识别率分别为73.725%、98.28%，优于SLPM、CL、JNS等人脸表情识别模型。

关键词: 卷积神经网络, 细粒度表情识别, 多尺度网络, 分层双线性池化, 多层特征融合

Abstract: Facial expressions are characterized by subtle differences between expression classes and significant changes within a class, which increases the difficulty of expression recognition.To address the problem, a neural network model is proposed based on multi-scale bilinear pooling.The global features of facial expressions are extracted by using three networks with different scales.Then a hierarchical bilinear pooling layer is introduced, and 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 ability of the model to represent and recognize subtle features of facial expressions. Multilayer feature information is fused by layer deconvolution, so the loss of key features that occurs when the neural network extracts features through multiple convolution layers and the pooling layer is solved.The experimental results show that the proposed model achieves a 73.725% recognition accuracy on FER2013 and 98.82% on CK+public data sets, outperforming SPLM, CL, JNS and other facial expression recognition algorithms.

Key words: Convolution Neural Network(CNN), fine-grained expression recognition, multiple-scale network, hierarchical bilinear pooling, multilayer feature fusion

中图分类号: 

• TP391