嵌入注意力的GaborCNN快速人脸表情识别方法

Gabor CNN Fast Facial Expression Recognition Method with Embedded Attention

人脸表情识别是智能人机交互研究的基础问题之一,面部情绪变化与嘴、眼睛、眉毛、鼻子等区域密切相关,这些特征对识别表情非常重要。为此,提出一个由4个Gabor滤波卷积层、注意力模块和两个全链接层组成的注意力Gabor卷积网络,同时使用不平衡损失focalloss对网络进行优化。首先,通过Gabor核与传统卷积滤波器调制的Gabor定向滤波器相较于传统卷积滤波器能更好地捕获感兴趣区域的信息,然后利用通道注意力、空间注意力模块提取区域中更关键的特征。在FERPlus和RAF-DB数据集上的实验表明,该模型结构简单、易于训练、计算成本低,识别精度分别达到88.39%、87.22%。

Facial expression recognition is one of issue in intelligent human-computer interaction research.Facial emotion changes are closely related to areas of interest such as the mouth,eyes,eyebrows,nose,etc.These features are very important for recognizing facial expressions.To this end,an attention Gabor convolutional network consisting of four Gabor filtering convolutional layers,an attention module,and two ful-ly linked layers is proposed,and the network is optimized using imbalanced loss focal loss.Firstly,Gabor directional filters modulated by Ga-bor kernels and traditional convolutional filters can better capture information about regions of interest compared to traditional convolutional fil-ters.Then,channel attention and spatial attention modules are used to extract more critical features in the region.The experiments on FER-Plus and RAF-DB datasets show that the model has a simple structure,is easy to train,and has low computational costs.The recognition accu-racy reaches 88.39%and 87.22%,respectively.