基于改进EfficientNet的表情识别方法
Facial Expression Recognition Method Based on Improved EfficientNet
丁祥 1唐宏伟 2石书琪 2高方坤 1罗佳强 1王军权1
作者信息
- 1. 邵阳学院机械与能源工程学院,湖南邵阳 422099
- 2. 邵阳学院电气工程学院,湖南邵阳 422099
- 折叠
摘要
针对现有网络模型表情特征提取不充分以及表情数据存在类内差距大、类间差距小的问题,对模型训练中全局特征和局部特征的提取方法展开了研究,并提出了基于改进EfficientNet的表情识别方法.首先,在浅层网络中使用了大核Fused-MBConv卷积块提取全局特征;然后,在深层网络中使用了小核MBConv卷积块提取局部特征,并结合ACON激活函数,以EfficientNetB0 为基线网络,构建了LA-EfficientNetB0 网络;最后,通过Grad-CAM显示不同模型提取的特征对原图的关注区域,论证了本文全局特征和局部特征提取方法的有效性.结果表明,LA-EfficientNetB0在FER2013人脸表情数据集准确率达71.61%,优于VGG16、ResNet50、EfficientNetB0、EfficientNetV2B0网络模型.
Abstract
In response to the insufficient extraction of expression features in existing network models and the problems of large intra class and small inter class differences in expression data,research was conducted on the extraction methods of global and local features in model training,and an expression recognition method based on improved EfficientNet has been proposed.Firstly,large kernel Fused-MBConv convolutional blocks were used in shallow networks to extract global features.Then,small kernel MBConv convolutional blocks were used in the deep network to extract local features,and combined with the ACON activation function,the LA-EfficientNetB0 network was constructed with EfficientNetB0 as the baseline network.Finally,the effectiveness of the global and local feature extraction methods proposed in this paper was demonstrated by displaying the regions of interest of the original image with the features extracted by different models through Grad-CAM.The results showed that the accuracy of LA-EfficientNetB0 in the FER2013 facial expression dataset reached 71.61%,which is better than the VGG16,ResNet50,EfficientNetB0,and EfficientNetV2B0 network models.
关键词
表情识别/EfficientNet/ACON/注意力机制Key words
expression recognition/EfficientNet/ACON/attention mechanism引用本文复制引用
基金项目
湖南省自科基金(2022JJ50205)
湖南省科技计划项目(2016TP1023)
湖南省科技计划项目(2023TP2036)
湖南省研究生科研创新项目(CX20221314)
邵阳学院研究生科研创新项目(CX2022SY005)
邵阳学院研究生科研创新项目(CX2022SY023)
出版年
2024
NETL
NSTL
万方数据