基于TPLBP/HOG特征与DBN深度模型的人脸识别研究

针对传统人脸识别方法所提取的人脸信息特征较为单一,且分类算法存在局限性的问题,在多特征信息融合的基础上结合深度信念网络(DBN)对人脸进行深度训练并进行识别。首先采取对比度受限自适应均衡化对人脸图像进行预处理,从而削弱光照对人脸识别的影响;然后,将提取到的人脸图像的TPLBP纹理特征和HOG结构特征进行特征融合,得到信息互补的融合特征;最后,将降维后的融合特征作为DBN的输入,通过对DBN深度模型的参数的动态搜索确定最佳值后,基于训练好的深度信念网络实现人脸图像样本的识别。以ORL、AR和Yale-B人脸数据库为基础进行试验,试验结果表明本文方法相较于传统的SVM、KNN和DBN算法准确率有很大提高,鲁棒性强。

Face Recognition Based on TPLBP/HOG Features and DBN Depth Model

The features of face information extracted by traditional face recognition methods are relatively single,and the classification algorithm has limitations.Combined with the deep belief network(DBN),the face is deeply trained and recognized based on multi-feature information fusion.Firstly,the face image is adaptively equalized,thereby the influence of illumination on face recognition is weakened.Then,TPLBP and HOG features that are robust to light,rotation are extracted and their feature information is effectively fused.Finally,based on multi-feature information after dimensionality reduction,a face recognition algorithm based on DBN was proposed after dynamically searching DBN depth model parameters to determine the optimal value.Experiments were conducted on the basis of ORL,AR,Yale-B face database.The experimental results show that compared with the traditional SVM,KNN and DBN algorithms,the accuracy of the proposed method is greatly improved,and the robustness is strong.