-
非约束环境下采集的人脸图像复杂多变, 将其直接作为字典原子用于稀疏表示分类(sparse representation based classification, SRC), 识别效果不理想. 针对该问题, 本文提出一种基于旋转主方向梯度直方图特征的判别稀疏图映射(discriminative sparse graph embedding based on histogram of rotated principal orientation gradients, DSGE-HRPOG)算法, 用于构建类内紧凑、类间分离的低维判别特征字典, 提高稀疏表示分类准确性. 首先, 采用旋转主方向梯度直方图(histogram of rotated principal orientation gradients, HRPOG)特征算子提取非约束人脸图像的多尺度多方向梯度特征, 有效去除外界干扰和像素间冗余信息, 构建稳定、鉴别的HRPOG特征字典; 其次, 引入判别稀疏图映射(discriminative sparse graph embedding, DSGE)算法, 以类内重构散度最小、类间重构散度最大为目标计算特征字典的最佳低维投影矩阵, 进一步增强低维特征字典的判别性、紧致性; 最后, 提出投影矩阵和稀疏重构关系交替迭代优化算法, 将维数约简过程伴随在稀疏图构建过程中, 使分类效果更理想. 在AR, Extended Yale B, LFW和PubFig这4个数据库上进行大量实验, 验证了本文算法在实验环境数据库和真实环境数据库上的有效性.The unconstrained face images collected in the real environments include many complicated and changeable interference factors, and sparsity preserving projections (SPP) cannot well obtain the low-dimensional intrinsic structure embedded in the high-dimensional samples, which is important for subsequent sparse representation classifier (SRC). To deal with this problem, in this paper we propose a new method named discriminative sparsity graph embedding based on histogram of rotated principal orientation gradients (DSGE-HRPOG). Firstly, it extracts multi-scale and multi-directional gradient features of unconstrained face images by HRPOG feature descriptor and incorporates them into a discriminative feature dictionary of sparse representation classifier. Secondly, it seeks an optimal subspace of HRPOG feature dictionary in which the atoms in intra-classes are as compact as possible, while the atoms in inter-classes are as separable as possible by adopting the proposed DSGE dimensionality reduction method. Finally, an optimal algorithm is presented in which the low-dimensional projection and the sparse graph construction are iteratively updated, and the accuracy of unconstrained face recognition is further improved. Extensive experimental results on AR, Extended Yale B, LFW and PubFig databases demonstrate the effectiveness of our proposed method.
-
Keywords:
- unconstrained face recognition/
- manifold learning/
- sparsity preserving projection/
- histogram of oriented gradients
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] -
Experiment 1/% Experiment 2/% Experiment 3/% LPP[22] 71.39 68.68 69.46 NPE[23] 72.64 71.81 71.08 SPP[33] 75.90 72.92 74.07 DSNPE[37] 80.28 78.26 78.14 SRC-DP[40] 78.35 76.50 77.80 SRC-FDC[42] 80.90 79.90 80.30 DSGE-pixels 79.03 78.75 82.65 DSGE-HRPOG
(3-HRPOG)88.54 89.51 90.53 DSGE-HRPOG
(5-HRPOG)89.31 89.58 90.98 DSGE-HRPOG
(Ms-HRPOG)89.31 90.00 91.06 
下载:
导出CSV
Experiment 1/% Experiment 2/% LPP[22] 95.51 ± 0.40 96.78 ± 0.72 NPE[23] 96.43 ± 0.23 97.85 ± 0.31 SPP[33] 92.57 ± 0.84 93.05 ± 0.77 DSNPE[37] 94.18 ± 0.48 95.29 ± 0.54 Gao[53] 86.91 ± 1.07 88.23 ± 0.91 DSGE-pixels 95.83 ± 0.66 96.21 ± 0.21 DSGE-HRPOG(3-HRPOG) 97.30 ± 0.20 97.73 ± 0.35 DSGE-HRPOG (5-HRPOG) 96.85 ± 0.38 96.93 ± 0.60 DSGE-HRPOG G(Ms-HRPOG) 97.98 ± 0.50 98.10 ± 0.31 下载:
导出CSV
LFW/% PubFig/% LPP[22] 35.32 24.00 NPE[23] 35.19 25.00 SPP[33] 31.52 29.00 DSNPE[37] 44.05 30.90 WGSC[58] 47.60 37.50 RSRC[3] 42.80 47.00 RRC[57] 53.20 42.20 IRGSC[41] 56.30 48.50 DSGE-pixels 51.52 38.60 DSGE-HOG 69.62 49.00 DSGE-HRPOG(3-HRPOG) 76.71 54.20 DSGE-HRPOG (5-HRPOG) 76.58 53.30 DSGE-HRPOG (Ms-HRPOG) 73.80 53.70 下载:
导出CSV
DSGE-HRPOG 3-HRPOG 5-HRPOG Ms-HRPOG with joint optimization 54.20 (630) 53.30 (473) 53.70 (514) without joint optimization 53.50 (514) 50.90 (423) 53.20 (514) 下载:
导出CSV
-
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60]
下载:
计量
- 文章访问数:7973
- PDF下载量:38
- 被引次数:0