基于HOG特征的加权稀疏表示人脸识别算法

Weighted Sparse Representation Face Recognition Algorithm Based on HOG Feature

在人脸识别中,为了进一步提高人脸图像对光照等外界因素的鲁棒性,提出一种基于HOG特征的加权稀疏表示算法,将加权稀疏表示方法和HOG特征以及随机投影方法相融合,以降低复杂度,提高识别性能。首先,统计每一幅图像的方向梯度直方图(HOG)特征,并对每一幅图像进行归一化处理,削弱人脸图像中的光照影响;其次,对归一化后的图像引入随机矩阵算法,进行多次随机投影,得到每个样本所对应的稀疏系数,利用样本之间的距离作为稀疏系数的权值;在此基础上,对传统稀疏表示分类器进行改进,样本经随机矩阵多次投影和稀疏表示后会产生多个重构残差,最后利用样本的重构残差和对样本进行识别分类。ORL人脸库和GT人脸数据库上的实验证明该方法对光照等外界物理因素有着很好的鲁棒性。

In face recognition,in order to further improve the robustness of face image to external factors such as illumination,a weighted sparse representation algorithm based on hog feature is proposed.The weighted sparse representation method is combined with HOG feature and random projection method to reduce complexity and improve recognition performance.Firstly,the HOG features of each image are counted,and each image is normalized to weaken the influence of illumination in the face image.Secondly,the random matrix algorithm is introduced into the normalized image,and the sparse coefficient corresponding to each sample is obtained by multiple random projections.The distance between samples is used as the weight of the sparse coefficient.On this basis,the traditional sparse representation classifier is improved.After multiple projections of random matrix and sparse representation,multiple reconstruction residuals will be generated.Finally,the reconstruction residuals of samples are used to identify and classify samples.Experiments on ORL face database and GT face database show that the proposed method is robust to external physical factors such as illumination.