Facial expression recognition based on Image Euclidean Distance-Supervised Neighborhood Preserving Embedding

摘要

High-dimensional data often lie on relatively low-dimensional manifold, while the nonlinear geometry of that manifold is often embedded in the similarities between the data points. These similar structures are captured by Neighborhood Preserving Embedding (NPE) effectively. But NPE as an unsupervised method can't utilize class information to guide the procedure of nonlinear dimensionality reduction. They ignore the geometrical structure information of local data points and the spatial information of pixels, which leads to the failure of classification. For this problem, a feature extraction method based on Image Euclidean Distance-Supervised NPE (IED-SNPE) is proposed, and is applied to facial expression recognition. Firstly, it employs Image Euclidean Distance (IED) to characterize the dissimilarity of data points. And then the neighborhood graph of the input data is constructed according to a certain kind of dissimilarity between data points. Finally, it fuses prior nonlinear facial expression manifold of facial expression images and class-label information to extract discriminative features for expression recognition. In the classification experiments on JAFFE facial expression database, IED-SNPE is used for feature extraction and compared with NPE, SNPE, and IED-NPE. The results reveal that IED-SNPE not only the local structure of expression manifold preserves well but also explicitly considers the spatial relationships among pixels in the images. So it excels NPE in feature extraction and is highly competitive with those well-known feature extraction methods.