基于MobileNetV2特征提取的自修复表情识别算法  

Self-Cure Expression Recognition Algorithm Based on MobileNetv2 Feature Extraction

作　　者：于延[1] 刘忠旭 YU Yan;LIU Zhong-xu(Harbin Normal University,School of Computer Science and Information Engineering,Harbin 150025,Heilongjiang)

机构地区：[1]哈尔滨师范大学计算机科学与信息工程学院,黑龙江哈尔滨150025

出　　处：《电脑与电信》2023年第9期78-82,共5页Computer & Telecommunication

摘　　要：为了减少自修复网络(Self-Cure Network,SCN)表情识别算法的计算量、参数量,提高模型精度,采用改进的MobileNetV2来做人脸表情图像的特征提取。调整MobileNetV2的网络层结构;在模型中嵌入CA注意力机制;采用全局逐深度卷积(GDConv)代替全局平均池化层;调整深度卷积的卷积核大小;调整宽度因子降低模型的参数量。改进后的算法在公共标准数据集RAF-DB和FERPlus上精度分别较SCN提高了7.81%和5.37%,且具备SCN的算法优点;参数量和计算量分别是原模型的1.5219%和4.9511%,且远低于其他对比模型,为轻量级表情识别算法的研究提供参考。In order to reduce the calculation amount and parameters of the Self-Cure Network(SCN) expression recognition algorithm and improve the accuracy of the model,the improved MobileNetV2 network is used to extract the features of facial expressions.In this paper,it adjusts the network structure of MobileNetV2,embeds CA attention mechanisms in the model,and uses Global depthwise convolution(GDConv) instead of the global average pooling layer.It adjusts the size of the convolution kernel of the deep convolution,and adjusts the width factor reducing the number of parameters for the model.The improved algorithm has 7.81%and 5.37% higher accuracy on the common standard data sets RAF-DB and FERPlus than SCN respectively,and has the advantages of SCN algorithm.The parameter quantity and calculation amount are 1.5219% and 4.9511% of the original model respectively,and are much lower than other comparison models,which provides a reference for the research of lightweight expression recognition algorithm.

关 键 词：人脸表情识别技术 MobileNetV2 注意力机制 特征提取 自修复网络 卷积神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]