半导体光电, 2015, 36 (5): 815, 网络出版: 2015-11-30
结合Gabor特征和深度信念网络的人脸姿态分类
Multipose Face Classification Based on the 2DGabor features and Deep Belief Nets
深度信念网络 姿态分类 BP算法 二维Gabor特征 近邻元分析 deep belief nets gesture classification backpropagation algorithm 2Dgabor feature neighborhood component analysis
摘要
针对多姿态人脸图像分类存在的困难,提出了一种基于Gabor特征和深度信念网络(DBN)的近邻元分析(NCA)方法,通过提取Gabor多姿态人脸图像的尺度图并将其进行融合,从而对多姿态人脸图像具有较好的区分度,利用融合后的特征图来训练样本并作为深度信念网络的输入图像,结合NCA分析对训练样本进行线性变化以寻找到一个更有利于类别分类的线性子空间,提供足够大的数据集来估算模型参数进而对多姿态人脸图像进行分类。对ORL人脸数据集测试结果表明,多姿态人脸分类数据量为1616和2432之间时的平均分类正确率分别为86.67%、84.00%、90.67%和86.67%,与PCA、LDA和RCA三种算法相比,其分类准确率都得到了提高,实验结果验证了这种针对多姿态人脸图像的分类算法的有效性。
Abstract
Aiming at the pose problems of face image which may severely degrade the classification performance,proposed was a multipose face classification method based on the 2DGabor features and deep belief nets (DBNs) approach to construct deep learning for classifying multipose images accurately. By extracting 2DGabor features of multipose faces and fusing them,it is aimed to improve the features learning with more discriminating power to benefit the classification problems. By using fused images as the input image in deep belief nets and incorporating the neighborhood component analysis,it is to change the linearly training samples so as to find a more favorable category of linear subspace,then we can provide a large enough data set can be provided to estimate model parameters. The average classification accuracies of the proposed algorithm on the ORL images datasets are 86.67%、84.00%、90.67% and 86.67% respectively when the multipose face classified data is between 16×16 and 24×32. The classification accuracy is improeved compared to the PCA、LDA and RCA methods. Experimental results verifies the effectiveness of proposede algorithm.
陈勇, 黄婷婷, 张开碧, 郝裕斌, 帅峰. 结合Gabor特征和深度信念网络的人脸姿态分类[J]. 半导体光电, 2015, 36(5): 815. CHEN Yong, HUANG Tingting, ZHANG Kaibi, HAO Yubin, SHUAI Feng. Multipose Face Classification Based on the 2DGabor features and Deep Belief Nets[J]. Semiconductor Optoelectronics, 2015, 36(5): 815.