基于卷积神经网络的人脸表情识别算法

Facial Expression Recognition Algorithm Based on Convolutional Neural Network

摘要: 人脸表情识别是图像识别的一个重要领域。传统人脸表情识别主要基于人工提取特征，由于人脸表情丰富、背景复杂、差异范围大等问题，其存在算法鲁棒性较差、易受人脸身份信息干扰等问题，以及传统卷积神经网络易发生过拟合、梯度弥散、梯度爆炸等问题的现状，因此本文提出一种多层特征融合非反向传播稠密卷积神经网络的人脸表情识别算法。该算法应用了改进的HSIC (Hilbert-Schmidt Independence Criterion，希尔伯特–施密特独立性)-bottleneck来代替传统反向传播(Back Propagation, BP)，具有诸多独特的优点。在特征提取过程中，为了充分利用得到的特征图像，将卷积层稠密连接并引入attention机制，最终通过softmax分类器分类，得到分类结果。在FER2013数据集上进行了多次实验，与传统BP算法的卷积神经网络算法相比，不仅有效减轻过拟合现象，并且在模型收敛速度上更快、计算量更小、内存占用更小，证明了在人脸表情识别问题中非反向传播稠密卷积神经网络模型结构有效、提出的分类优化方法可行。

Abstract: Facial expression recognition is an important field of image recognition. Traditional facial expression recognition is mainly based on manual extraction of features, which has the problems of poor algorithm robustness and susceptibility to interference by face identity information due to rich facial expressions, complex background and large range of differences, as well as the current situation that traditional convolutional neural networks are prone to overfitting, gradient dispersion and gradient explosion, etc. Therefore, this paper proposes a multilayer feature fusion using dense convolutional neural network without Back Propagation (BP) for face expression recognition algorithm. The algorithm applies a modified HSIC (Hilbert-Schmidt independence criterion)-bottleneck instead of the traditional BP, which has many unique advantages. In the feature extraction process, in order to make full use of the obtained feature images, the convolutional layers are densely connected and attention mechanism is introduced, and finally the classification results are obtained by the softmax classifier. Compared with the traditional BP algorithm of convolutional neural network algorithm, it not only effectively reduces the overfitting phenomenon, but also has faster model convergence, smaller computation and smaller memory consumption, which proves that the structure of non-back propagation dense convolutional neural network model is effective and the proposed classification optimization method is feasible in the face expression recognition problem.