改进的HOG和Gabor,LBP性能比较

Performance Comparison of Improved HOG,Gabor and LBP

为了实现复杂环境下的人脸特征有效表达,提出一种改进的梯度方向直方图(HOG)人脸识别方法.首先以人脸图像网格作为采样窗口并在其上提取HOG特征;然后将所有网格HOG特征向量进行组合,实现整个人脸特征表达;最后采用最近邻分类器进行识别.另外,比较了该方法与Gabor小波和局部二值模式(LBP)2种著名的人脸局部特征表示方法的优劣.实验结果表明,在调优的HOG参数下,在具有光照和时间环境等复杂变化的FERET人脸库中,较少维数的HOG特征比LBP特征有更好的表现,而且HOG特征提取时间和特征向量维数比Gabor小波方法更具有优势.

In order to achieve effective expression of the facial features in complex environment, an improved face recognition method of histograms of oriented gradients （HOG） is proposed. Firstly the face image grid is set as a sampling window, in which HOG features are extracted. Then all grid HOG feature vectors are composed to realize the whole facial features expression. Finally, recognition is achieved with a nearest neighbor classifier. We further compare it with two popular face local features expression methods, Gabor wavelet and local binary pattern （LBP）. With the optimistic HOG parameters, experimental results show that HOG features with less dimensions have better performance than the LBP features in the FERET face collection under complex changes of light and time environments. Meanwhile HOG features have advantages over Gabor wavelet with less computational time of feature extraction and smaller number of feature vector dimensions.