基于多级纹理特征的深度信念网络人脸识别算法

为解决人脸特征提取过程中局部特征缺失的问题,借助局部二值模式(LBP)与方向梯度直方图(HOG)提出一种基于多级纹理特征融合的深度信念网络人脸识别算法。以提取局部纹理特征以及边缘纹理特征为出发点,对人脸图像进行三级纹理特征提取。使用MB-LBP提取初级纹理特征;在此基础上进行改进的CS-LBP图像特征提取作为二级纹理特征;使用HOG算子在二级纹理特征上完成三级纹理特征提取。将二级和三级纹理特征直方图顺序串联融合后输入到深度信念网络(DBN)逐层贪婪训练,优化网络参数,并用优化的网络在ORL、YELA人脸标准库中进行测试,识别率均在92%以上。该算法与传统算法(SVM、PCA)相比较拥有更好的人脸识别效果,同时也表明了局部纹理特征的改善为识别过程的特征提取提供强有力的保障,为人脸识别的进一步研究开拓新思路。

DEEP BELIEF NETWORK FACE RECOGNITION ALGORITHM BASED ON MULTI-LEVEL TEXTURE FEATURES

In order to solve the problem of local feature missing in the process of face feature extraction,we propose a deep belief network face recognition algorithm based on multi-level texture feature by means of local binary pattern(LBP)and histograms of oriented gradients(HOG).It starts with the extraction of local texture features and edge texture features,and performs three-level texture feature extraction for face image.We used MB-LBP to extract the primary texture features;on this basis,improved CS-LBP image features were extracted as the secondary texture features;the HOG operator was used to complete the extraction of the third-level texture features on the secondary texture features.Subsequently,the second and third order texture feature histograms were concatenated and input into the deep belief network(DBN)layer by layer greedy training to optimize the network parameters.The optimized network was tested in ORL and YELA face standard database,and the recognition rate was over 92%.Compared with traditional algorithms(SVM,PCA),our algorithm has a better face recognition effect.It also shows that the improvement of local texture features provides a strong guarantee for feature extraction in the recognition process,and opens up new ideas for further research on face recognition.