结合注意力机制的轻量化人脸表情识别方法

A Lightweight Facial Expression Recognition MethodCombining Attention Mechanism

针对目前人脸表情识别方法存在参数量多、计算资源消耗大和识别精度低的问题,研究一种轻量化的人脸表情识别方法。首先,对MobileNet V3L网络层数进行缩减,同时将倒残差结构中间通道数和输出通道数增大至原来的1.5~3.2倍。其次,引入改进的条件坐标注意力机制,在坐标信息嵌入中根据中间通道的数量,选择平均池化或最大池化进行编码,提取面部表情在空间和通道位置上的详细信息。最后,使用Mish代替h-swish激活函数,实现特征提取后的非线性化。在公开数据集FERPlus和RAF-DB进行实验,结果表明,所提方法比原始MobileNet系列模型识别精度分别提高0.60%和1.07%以上。所提方法比Ada-CM网络推理速度提升21.94%,识别精度提高0.49%以上,实验表明该方法具有良好的识别性能。

To address the problems of large number of parameters,high computational resource consumption and low recognition accuracy of current facial expression recognition methods,a lightweight facial expression recognition method is studied.Firstly,the number of layers of MobileNet V3L network is reduced,and the number of intermediate channels and output channels of inverse residual structure is increased to 1.5~3.2 times of the original one.Secondly,an improved conditional coordinate attention mechanism is introduced to extract detailed information of facial expressions in space and channel locations by choosing average pooling or maximum pooling for encoding in the coordinate information embedding according to the number of intermediate channels.Finally,Mish is used instead of the h-swish activation function to achieve nonlinearization after feature extraction.Experiments are conducted on the publicly available datasets FERPlus and PAF-DB,and the results show that the proposed method improves the recognition accuracy by more than 0.60%and 1.07%,respectively,over the original MobileNet series models.The proposed method improves the inference speed by 21.94%and the recognition accuracy by more than 0.49%over the Ada-CM network,and the experiments show that the method has good recognition performance.