作者:宋文博，高璐，苗壮，林克正
Authors:SONG Wen bo,GAO Lu,MIAO Zhuang,LIN Ke zheng
摘要:针对人脸表情网络模型参数复杂和计算性能偏低等问题 ,提出一种基于卷积注意力机制和深度可分离卷积 (convolutional block attention module-depthwise separable convolution, CBAM-DSC) 网络的表情识别方法 。网络使用深度可分离卷积与传统卷积相结合 ,提出的改进的Inception 模块通过不同分支提取不同特征信息的同时减少了网络参数量 ,提高网络运行的效率 。最后添加了卷积注意力机制模块 ,能够使网络提取特征时重点关注关键信息 ,从而使得提取到的特征表达信息更准确 ,更利于分类 。在 RAF-DB 数据集和 CK + 数据集上的仿真实验表明 ,网络模型 CBAM-DSC 具有较高的识别率 ,消融实验中相比传统 CNN(convolutional neural network) 网络的参数量减少了 6. 4% ,提升了计算性能。
Abstract:Aiming at the problems of complex parameters and low computational performance of facial expression network model, an expression recognition method based on Convolutional Block Attention Module-Depthwise Separable Convolution network is proposed. The network usage depth separable convolution is combined with the traditional convolution. The improved Inception module extracts different feature information through different branches while reducing the network parameters and improving the network operation efficiency. Finally, the convolutional block attention module is added, which can make the network focus on the key information when extracting features, so that the extracted feature expression information is more accurate and more conducive to classification. The simulation experiments on RAF-DB dataset and CK + dataset show that the network model Convolutional Block Attention Module-Depthwise Separable Convolution has a high recognition rate, and the number of parameters in the ablation experiment is reduced by 6. 4% compared with the traditional Convolutional Neural Network, which improves the computing performance.

PDF全文下载地址:

可免费Download/下载PDF全文