基于改进GD-HASLR算法的遮挡人脸识别

doi:10.16180/j.cnki.issn1007-7820.2023.06.011

摘要/Abstract

摘要：

针对遮挡人脸识别方面的算法在训练样本数目减少时,其识别结果也会下降。为了解决该问题,文中提出了一种改进的GD-HASLR(Gradient Direction-Based Hierarchical Adaptive Sparse and Low-Rank)算法。该算法先求得人脸图像的广义梯度方向,计算人脸图像从一阶到三阶的梯度大小和梯度方向,再利用映射函数进行映射后求得梯度方向向量,然后将其作为层次稀疏低秩模型的输入,求解出图像的表示系数和误差。文中采用了重启的快速迭代收缩阈值算法-II求解稀疏表示系数。最后,计算一阶到三阶测试样本的残差,选取其频率最高或者平均等级最低的类别作为分类结果。在AR、Extended Yale B数据库上的实验结果表明,与GD-HASLR等方法相比,文中改进方法获得的识别效果更好。

关键词: 遮挡, 人脸识别, 广义梯度方向, 梯度大小, 梯度方向, 层次稀疏低秩模型, 重启快速迭代收缩阈值算法-II, GD-HASLR

Abstract:

When the number of training samples is reduced, the recognition result of occlusion face recognition algorithm will also decrease. To solve this problem, an improved GD-HASLR algorithm is proposed. Firstly, the algorithm obtains the generalized gradient direction of the face image, and calculates the gradient size and gradient direction of the face image from the first order to the third order. Then, after mapping with the mapping function, the gradient direction vector is obtained, which is used as the input of the hierarchical sparse low-rank model, and the representation coefficient and error of the image are obtained. In this study, a restarted fast algorithm with shrinkage threshold-II is adopted to solve the sparse representation coefficient. Finally, the residuals of the first order to the third order test samples are calculated respectively, and the category with the highest frequency or the lowest average grade is selected as the classification result. Experimental results on AR and Extended Yale B databases show that compared with GD-HASLR and other methods, the recognition effect of the improved method proposed in this study is better.

Key words: occlusion, face recognition, generalized gradient direction, gradient size, gradient direction, hierarchical sparse low-rank model, restart fast iterative shrinkage thresholding algorithm-II, GD-HASLR

中图分类号:

TP391.4

徐恬恬,席志红. 基于改进GD-HASLR算法的遮挡人脸识别[J]. 电子科技, 2023, 36(6): 72-79.

XU Tiantian,XI Zhihong. Face Recognition with Occlusion Based on Improved GD-HASLR Algorithm[J]. Electronic Science and Technology, 2023, 36(6): 72-79.

图/表 17

图1

图2

图3

算法1 重启FISTA-II的计算过程

步骤1 给定初始值l₀,L=L(f),c₁=1,令h₁=l₀;
步骤2 while(当条件不收敛时,进行第t次迭代)
1)计算l_t,l_t=

S ξ α

(h_t-2αD^T(Dh_t-b))。其中,α=1/L(f),L(f)=2λ_max(D^TD),λ_max(·)表示矩阵的最大特征值,ξ为分层模型的自适应权重函数;ξ=y-

e k + 1

+(1/β)z^k ;
2)更新c_t₊₁,c_t₊₁=(1+sqrt(1+4

c t 2

))/2
if(重启条件成立)
计算h_t₊₁,h_t₊₁=l_t
else
计算h_t₊₁,h_t₊₁=l_t+((c_t-1)/c_t₊₁)(l_t-l_t_-1)
end while

图4

表1

表2

表3

表4

表5

表6

表7

表8

表9

表10

表11

表12

参考文献 20

[1]	刘满良, 孟辉, 王新宇, 等. 视频侦查中人脸识别准确率的影响因素分析[J]. 中国刑警学院学报, 2019(1):119-123.
	Liu Manliang, Meng Hui, Wang Xinyu, et al. Analysis of influencing on the accuracy of face recognition in video investigations[J]. Journal of China Criminal Police University, 2019(1):119-123.
[2]	彭荣杰, 彭亚雄, 陆安江. 基于改进PCA+SVM的人脸识别系统[J]. 电子科技, 2021, 34(12):56-61.
	Peng Rongjie, Peng Yaxiong, Lu Anjiang. Face recognition system based on improved PCA+SVM[J]. Electronic Science and Technology, 2021, 34(12):56-61.
[3]	李晓彤. 基于图像修复的局部遮挡人脸识别方法研究[D]. 哈尔滨: 哈尔滨理工大学, 2021:20-48.
	Li Xiaotong. Research on partial occlusion face recognition method based on image inpainting[D]. Harbin:Harbin University of Science and Technology, 2021:20-48.
[4]	吕世林. 结合概率协同字典学习和遮挡定位的非配合人脸识别[D]. 常州: 常州大学, 2021:30-56.
	Lü Shilin. Non-cooperative face recognition combining probabilistic collaborative dictionary learning and occlusion positioning[D]. Changzhou: Changzhou University, 2021:30-56.
[5]	Yang J, Luo L, Qian J, et al. Nuclear norm based matrix regression with applications to face recognition with occlusion and illumination changes[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(1):156-171. pmid: 26930675
[6]	Mishra G, Vishwakarma V P. Partial occlusion handling using quadrant-wise sparsity for face image recognition[C]. New Delhi: Proceedings of the Seventh International Conference on Computing for Sustainable Global Development, 2020:35-38.
[7]	Zhang J X, Liu H R. Local occluded face recognition based on 2D-DWT and sparse representation[C].Harbin:Proceedings of the Fifth International Conference on Mechanical, Control and Computer Engineering, 2020:2110-2114.
[8]	Du L, Hu H. Nuclear norm based adapted occlusion dictionary learning for face recognition with occlusion and illumination changes[J]. Neurocomputing, 2019, 34(2):133-144.
[9]	何惠慧. 基于核范数的有遮挡图像分类[D]. 西安: 西安电子科技大学, 2020:35-54.
	He Huihui. Image classification with occlusion based on kernel norm[D]. Xi'an: Xidian University, 2020:35-54.
[10]	凡正军. 噪声和遮挡条件下的人脸识别算法研究[D]. 郑州: 中原工学院, 2021:23-46.
	Fan Zhengjun. Research on face recognition algorithm under noise and occlusion conditions[D]. Zhengzhou: Zhongyuan Institute of Technology, 2021:23-46.
[11]	Lai S C, Kong M, Lam K M, et al. High-resolution face recognition via deep pore-feature matching[C]. Taipei: Proceedings of the IEEE International Conference on Image Processing, 2019:3477-3481.
[12]	Ge S, Li C, Zhao S, et al. Occluded face recognition in the wild by identity-diversity inpainting[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2020, 30(10):3387-3397. doi: 10.1109/TCSVT.76
[13]	Dong J, Zhang L, Zhang H, et al. Occlusion-aware gan for face de-occlusion in the wild[C]. London: Proceedings of the IEEE International Conference on Multimedia and Expo, 2020:1-6.
[14]	Wu C Y, Ding J J. Occluded face recognition using low-rank regression with generalized gradient direction[J]. Pattern Recognition, 2018, 80(6):256-268. doi: 10.1016/j.patcog.2018.03.016
[15]	李启朋. 非光滑凸优化问题的快速迭代收缩阈值算法研究[D]. 西安: 西安电子科技大学, 2018:35-65.
	Li Qipeng. Research on fast iterative shrinkage threshold algorithm for nonsmooth convex optimization problem[D]. Xi'an: Xidian University, 2018:35-65.
[16]	吴丽琼. 基于梯度的图像插值放大算法研究[D]. 济南: 山东大学, 2017:44-65.
	Wu Liqiong. Research on gradient-based image interpolation and enlargement algorithm[D]. Jinan: Shandong University, 2017:44-65.
[17]	Lee A, Caron F, Doucet A, et al. A hierarchical bayesian framework for constructing sparsity-inducing priors[M]. Paris:HAL-INRIA, 2010:59-77.
[18]	Murphy K P. Machine learning: A probabilistic perspective[M]. Cambridge: MIT Press, 2012:35-38.
[19]	侯小秋. 增广拉格朗日乘子法非严格成立的理论分析[J]. 中央民族大学学报(自然科学版), 2021, 30(3):9-15.
	Hou Xiaoqiu. Theoretical analysis of the non-strict establishment of the augmented Lagrangian multiplier method[J]. Journal of the Central University for Nationalities (Natural Science Edition), 2021, 30(3):9-15.
[20]	虞涛. 基于低秩稀疏分解的遮挡人脸识别研究[D]. 南京: 南京邮电大学, 2019:20-30.
	Yu Tao. Research on occluded face recognition based on low-rank sparse decomposition[D]. Nanjing: Nanjing University of Posts and Telecommunications, 2019:20-30.

方法	NMR/%	GD-HASLR/%	改进的GD-HASLR/%
实验1	63.33	75.33	88.33
实验2	37.67	50.33	74.00
实验3	56.33	70.67	75.00
实验4	35.00	50.67	60.00
实验5	46.92	61.75	74.33
实验6	60.67	82.67	94.33
实验7	61.00	82.00	94.67
实验8	63.33	73.33	80.67
实验9	62.67	73.00	79.67
实验10	60.92	77.75	87.33

方法	GD-HASLR/%	改进的GD-HASLR/%
sigmoid	75.33	88.33
tan^-1(x)	74.67	88.00
softsign	74.00	88.00
tanh(x)	77.67	88.00

方法	GD-HASLR/%	改进的GD-HASLR
sigmoid	50.33	74.00
tan^-1(x)	53.00	73.00
softsign	53.00	73.67
tanh(x)	53.33	73.67

方法	GD-HASLR/%	改进的GD-HASLR/%
sigmoid	70.67	75.00
tan^-1(x)	71.67	73.67
softsign	70.67	73.33
tanh(x)	71.67	76.00

方法	GD-HASLR/%	改进的GD-HASLR/%
sigmoid	50.67	60.00
tan^-1(x)	51.00	53.33
softsign	48.67	53.33
tanh(x)	53.00	54.67