基于Semi-Supervised LLE的人脸表情识别方法

Facial Expression Recognition Based on Semi-Supervised Local Linear Embedding

目的为提取有效的鉴别特征和降低鉴别向量的维数来识别人脸表情图像.方法将流行学习(Manifold learning,ML)和半监督学习(Semi-Supervised learning,SSL)结合起来,利用人脸表情图像数据本身的非线性流形结构信息和部分标签信息来调整点与点之间的距离形成距离矩阵,而后基于被调整的距离矩阵进行线性近邻重建来实现维数约简,提取低维鉴别特征用于人脸表情识别.结果该方法能充分利用数据的结构信息和有限的标签信息,使具有标签信息的同类样本之间的距离最小化,不同类数据之间的距离最大化,进而可以有效地提取数据的低维鉴别子流形,使得分类性能要优于非监督的维数约简方法.结论笔者提出的半监督局部线性嵌入算法能有效地提高人脸表情识别的性能.

In order to incorporate semi-supervised learning and manifold learning, especially to recognize facial expression, a new semi-supervised manifold learning for facial expression recognition is proposed. This method relied on the distance matrix formed by both labeled and unlabeled samples, and then the local linear embedding （LLE） method was used to extract discriminative manifold features according to the modified distance matrix. The proposed method produces better classification performance which captures the intrinsic manifold structure collectively revealed by labeled and unlabeled samples. Experimental results on public facial expression databases show that the proposed method can improves facial expression classification performance effectively.