基于改进VGG网络的面部表情识别方法

Facial Expression Recognition Method Based on Improved VGG Network

针对传统面部识别方法中网络模型重要通道特征关注度欠缺、参数过多、准确率低等问题,提出一种基于改进VGG19网络(Visual Geometry Group,VGG19)的表情识别方法。该方法在VGG19网络的每组卷积层之间都加入一个新模块,新模块由三方面组成:SE注意力机制模块、BN批量归一化层、PReLU激活函数,SE模块中的原激活函数ReLU替换为Mish激活函数,加速收敛,提升网络对面部细节关注度;对全连接层参数量进行修改,去除第一层全连接层和第二层全连接层,最大池化层替换为全局混合池化,达到卷积层加全局混合池化、一层全连接层的组合。原网络中全连接层节点数由[4096,4096,1000]改进为[512,7],改善了VGG网络庞大参数量的特征,增加抗过拟合效果。在CK+和FER-2013表情数据集上准确率分别达到98.990%和73.112%,证明所提方法具有较好的泛化性和准确率。

Targeting at the issues of insufficient attention to important channel features,excessive parameters,and low accuracy of tradi-tional facial recognition methods,an expression recognition method based on an improved Visual Geometry Group 19(VGG19)network is proposed.This method is fultilled by adding a new module between each group of convolution layers of VGG19 network.The new module is composed of three aspects,i.e.,SE attention mechanism module,BN batch normalization layer,and PReLU activation function.The original activation function ReLU in SE module is replaced by Mish activation function,which speeds up convergence and improves the attention to details on the surface of the network.The parameters of the full connection layer are modified,the full connec-tion layer of the first layer and the full connection layer of the second layer are removed,and the maximum pooling layer is replaced with the global mixed pooling,so as to achieve the combination of the convolution layer and the global mixed pooling,and the full connection layer of the first layer.The number of nodes of the full connection layer in the original network is changed from[4096,4096,1000]to[512,7],greatly reducing the disadvantage of large VGG network parameters,and increasing the anti overfitting effect.The accuracy rates on the CK+and FER-2013 facial expression datasets reach 98.990%and 73.112%,respectively,showing good generalization and accuracy of the proposed method.