基于L1范式的分块2DPCA人脸识别方法

摘要

在人脸识别中,增强人脸图像的重构效果和识别方法的鲁棒性一直是其中的技术难点.为了提高识别性能,先对图像矩阵进行分块,同时用一种新的图像信息熵自适应加权模式对人脸不同分块区域赋予不同的权值,然后直接应用L1范式代替L2范式进行图像特征抽取,最后用最近邻分类器进行分类.实验结果表明,新方法在识别性能上优于基于L1范式的2DPCA方法(2DPCA-L1),比2DPCA-L1更具有鲁棒性,显著地提高了有遮挡图像的重构效果.%It is a technical difficulty to enhance the reconstruction of face image and the robustness of recognition method. In order to improve recognition performance, the original images are divided into block images in the proposed approach. At the same time a new image entropy model of adaptive weighted is adopted for giving different block regions with different weights. Then L1-norm, instead of L2-norm is applied directly for image feature extraction. Finally, a nearest neighbor classifier is used for classification. The experimental results of ORL face database indicate that the recognition performance of new algorithm is better and robuster than two-dimensional principal component analysis based on Ll-norm method (2DPCA-L1). Because weighted models for better local feature extraction and robustness of Ll-norm are utilized, the effects of occluded image reconstruction are significantly improved.