| 摘要: |
| 目的 针对传统卷积神经网络在人脸表情识别过程中存在网络结构复杂、特征提取弱化更能反映情感状态
的眼睛和嘴唇区域而造成的针对性不强、忽略了人脸表情中空间结构信息导致识别准确度不高等问题,在主流人
脸表情识别方法的基础上提出了一种嵌入注意力机制和通道重排的人脸表情识别方法。 方法 首先,将预处理的人
脸图片传入空间注意力模块,通过增强空间维度信息,使得模型能够更好地关注图像中的关键区域;然后对获取到
的空间注意力特征图的通道进行切分,使信息流通过不同的路径,再进行通道融合,从而增强了特征表达能力;其
次,将得到的特征图传入到通道注意力模块,增强通道维度信息;最后,通过全局平均池化来进行网络预测。 结果
所设计的网络仅以 1. 9 M 的参数量在数据集 FER2013 和 CK+上分别达到 71. 80%和 99. 66%的识别准确度。 结论
该方法相比许多传统经典算法有更好的识别效果,为提高人脸表情识别准确度提供了一种更有效的途径,具有很
好的实用价值和应用前景。 |
| 关键词: 人脸表情识别 残差网络 通道重排 注意力机制 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Face Expression Recognition with Embedded Attention Mechanism and Channel Shuffle |
|
WANG Zhengquan ZHU Chengjie
|
|
School of Electrical and Information Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective Traditional convolutional neural networks face challenges in face expression recognition due to
complex network structures and weakened feature extraction particularly from critical regions like eyes and lips that reflect
emotional states. This leads to issues such as weak specificity and neglect of spatial structural information in facial
expressions resulting in lower recognition accuracy. In response this study proposes a face expression recognition method
embedded with attention mechanisms and channel shuffle based on mainstream approaches in the field. Methods Firstly
preprocessed facial images are fed into a spatial attention module to enhance spatial dimensional information allowing the
model to better focus on key areas within the images. Secondly the channels of the obtained spatial attention feature map
are split to make the information flow pass through different paths and then channel fusion is carried out to enhance the
ability of feature expression. Subsequently the obtained feature maps are introduced into the channel attention module to
enhance the channel dimension information. Finally global average pooling is applied for network prediction. Results
The designed network achieves recognition accuracies of 71. 80% on the FER2013 dataset and 99. 66% on CK+ with only
1. 9 M parameters. Conclusion This method outperforms many traditional classic algorithms providing a more effective
approach to improving face expression recognition accuracy. This method holds significant practical value and promising
application prospects. |
| Key words: face expression recognition residual networks channel shuffle attention mechanism |
