基于有监督降维的人脸识别方法

Face Recognition Method Based on Supervised Dimensionality Reduction

局部降维方法中存在仅考虑图像的相似信息,不能较好地保持图像的差异信息和像素间的空间结构等问题。为此,提出一种新的有监督降维方法,通过构建局部邻域相似图和局部差异图来刻画图像的局部结构。考虑到像素的空间结构,引入二维离散拉普拉斯图的光滑正则化来约束变换矩阵的平滑性。在Yale和ORL人脸数据库上进行实验验证,结果表明,该降维方法既能保持图像之间的局部结构信息,又能较好地保持图像间的差异信息及像素间的空间结构,并针对人脸图像可以有效提取出具有区分能力的低维特征,具有较高的识别精度。

Traditional dimensionality reduction methods only pay attention to the local similarity information of images. They neglect the diversity information of images and spatial structure of the pixels in the images. Therefore, a new supervised dimensionality reduction method is proposed, which constructs the local similarity graph and local diversity graph to characterize the local structure of images. Furthermore, a 2D Discretized Laplacian Smooth regularization by exploiting the spatial structure of the pixels in the images is introduced into the objective function. The method effectively maintains the local structure information between images and maintains the diversity information between images and spatial structure of the pixels in the images. It can effectively extract out the low dimensional feature from the face image. The method is verified on the Yale and ORL database, and experimental results show that the method has high recognition accuracy.