| 李波,王坤侠.基于特征融合的人脸表情识别算法研究[J].安徽建筑大学学报,2021,29(1):94-102 |
| 基于特征融合的人脸表情识别算法研究 |
| Facial Expression Recognition Method Based on Feature Fusion |
| |
| DOI： |
| 中文关键词: 人脸表情识别 方向梯度直方图 人脸特征点 特征融合 |
| 英文关键词: facial expression recognition histogram of oriented gradient facial landmark feature fusion |
| 基金项目:安徽省自然科学基金面上项目（1708085MF167） |
|
| 摘要点击次数: 6370 |
| 全文下载次数: 5774 |
| 中文摘要: |
| 为了避免传统表情识别方法中复杂的特征手动提取过程，同时保证能够提取到更多的表情特征，文
中提出一种融合卷积神经网络（Convolutional Neural Network，CNN）、方向梯度直方图（Histogram of Oriented
Gradient，HOG）以及人脸关键点定位（facial landmark detection）的人脸表情识别方法。该方法首先通过在图像
预处理中使用多任务卷积神经网络（Multi-task convolutional neural network，MTCNN）对不同尺度输入图像进行
人脸检测并得到人脸的关键点位置信息（facial landmark）。然后根据 facial landmark 的位置信息计算出人脸表
情图像的几何结构特征，并且计算人脸表情图像局部区域的方向梯度直方图来构成 HOG 特征，采用特征融合
的方式将 facial landmark 和 HOG 特征做进一步的融合形成新的特征向量 LM_HOG。最后将融合后的特征与经过
CNN 提取的全局特征再次融合输入到支持向量机（Support Vector Machine，SVM）和 Softmax 中进行表情识别。
在 FER2013 和 Extended Cohn-Kanade（CK+）人脸表情库上的实验结果表明，将融合得到的 LM_HOG 特征作为
局部特征，用以描述图像的局部性差异，CNN 提取的特征作为全局特征，用以描述人脸表情图像的整体性差异，
融合后的特征能更好的提取图像细节特征，平均识别率分别达到了 75.14% 和 97.86%，具有优越的性能。 |
| 英文摘要: |
| In order to avoid the complex manual feature extraction process in traditional facial expression recognition methods and
to extract more facial features，this paper proposes a facial expression recognition method based on convolutional neural network，
histogram of oriented gradient and facial landmark detection. First，the method uses multi-task convolutional neural network to perform
face detection on input images of different scales and obtain landmark position information of the face in image preprocessing. Then，
calculate the geometric structural features of the facial expression image according to the location information of the facial landmark，
and calculate the histogram of oriented gradient of the local area of the facial expression image to form the HOG feature. The feature
fusion method is used to make the facial landmark and the HOG feature forms a new feature vector LM_HOG. Finally，the fused features
and global features extracted by CNN are fused again and input into the SVM and Softmax for expression recognition. The experimental
results on the FER2013 and Extended Cohn-Kanade （CK+） show that the LM_HOG features obtained by the fusion are used as local
features to describe the local differences of the images，and the features extracted by CNN are used as global features to describe the
overall difference in facial expression images，the fused features can better extract the image details，and the average recognition rate
has reached 75.14% and 97.86%，respectively，and has superior performance |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
