自适应多目标遗传算法的集成剪枝用于人脸表情识别

doi:10.16180/j.cnki.issn1007-7820.2023.12.008

摘要/Abstract

摘要：

在集成剪枝中,为了同时高效地选择优质、独立的分类器,文中提出了一种新的动态自适应交叉策略的遗传算法用于分类器的集成剪枝。该方法使用轮盘赌和贪婪策略动态更新每个交叉策略的优先级,根据优先级计算每个策略被选中的概率,从而在算法迭代过程中自适应选择不同的交叉策略。此外,该方法考虑了交叉概率和变异概率动态自适应变化,并使用大多数投票法对挑选出来的分类器进行集成以获得最终结果。将文中所提方法与一些集成剪枝方法在5个真实人脸表情数据集上进行对比,实验结果表明文中所提该方法可以选出效果更好、冗余度更低的分类器,并在CK+数据集上具有22.50%的最低误差。

关键词: 人脸表情识别, 集成剪枝, 多目标遗传算法, 轮盘赌, 自适应交叉策略, 动态交叉概率, 动态突变概率, 大多数投票

Abstract:

In ensemble pruning,a new genetic algorithm with dynamically adaptive crossover strategies is proposed for the ensemble pruning of classifiers to simultaneously and efficiently select high-quality,independent classifiers.The method dynamically updates the priority of each crossover strategy using roulette wheel blocking and greedy strategies,calculates the probability of each strategy being selected based on the priority, and thus adaptively selecting different crossover strategies during the iteration of the algorithm.The method also considers the dynamic adaptive change of crossover probability and mutation probability and integrates the selected classifiers using majority voting to obtain the final result.The proposed method in the study is compared with some ensemble pruning methods on five real face expression data sets.The experimental results show that the proposed method can select classifiers with better results and lower redundancy,and it has the lowest error of 22.50% on the CK+ data set.

Key words: face expression recognition, ensemble pruning, multi-objective genetic algorithm, roulette wheel blocking, adaptive crossover strategy, dynamic crossover probability, dynamic mutation probability, majority voting

中图分类号:

陈星,李丹杨,何庆. 自适应多目标遗传算法的集成剪枝用于人脸表情识别[J]. 电子科技, 2023, 36(12): 55-63.

CHEN Xing,LI Danyang,HE Qing. Adaptive Multi-Objective Genetic Algorithm with Ensemble Pruning for Facial Expression Recognition[J]. Electronic Science and Technology, 2023, 36(12): 55-63.

图/表 14

图1

图2

图3

表1

图4

图5

算法2 AMGAEP

输入:训练集T={(x_i,x_j)

} i = 1 n

,验证集V={(x_i,x_j)

} i = 1 n

,测试集T={(x_i,x_j)

} i = 1 n

。
输出:集成剪枝错误率err,最优的分类器子集cls={c_i(x)}

i = 1 k

。
1.使用训练集训练一组基分类器C=

{c i (x)} i = 1 m

2.初始化:种群大小popsize,交叉策略S_m=1,m∈{1,2,3,4},当前迭代次数t,最大迭代次数maxIter,δ和c等
3.通过式(1)在验证集V中获取分类器预测值
4.通过式(2)和式(3)计算种群的适应度值
5.非支配排序
6.计算拥塞距离
7.While t<maxIter
8. 锦标赛选择子代
9. 根据式(7)确定交叉子代的数量
10. 使用算法1进行交叉操作生成个子代
11. 对种群进行突变
12. 使用式(4)更新突变概率
13. 父代和子代合并
14. 非支配排序
15. 计算拥塞距离
16. 精英策略选择
17. 更新帕累托前沿PF
18. end while
19. PF的所有染色体在测试集中测试最终集成剪枝的效果,选择具有最小错误率的染色体,并输出错误率err和分类器子集cls

表2

图6

图7

表3

表4

图8

参考文献 22

[1]	Li D, Wen G, Li X, et al. Graph-based dynamic ensemble pruning for facial expression recognition[J]. Applied Intelligence, 2019, 49(9):3188-3206. doi: 10.1007/s10489-019-01435-2
[2]	Saxena S, Tripathi S, Sudarshan T. An intelligent facial expression recognition system with emotion intensity classification[J]. Cognitive Systems Research, 2022, 74(4):39-52. doi: 10.1016/j.cogsys.2022.04.001
[3]	Hakim A, Marsland S, Guesgen H W. Computational representation and analysis of emotion dynamics[J]. Multimedia Tools and Applications, 2022, 81(15):21111-21133. doi: 10.1007/s11042-022-12490-2
[4]	Yca B, Sea C, Koa D, et al. Emotional context effect onrecognition of varying facial emotion expression intensities in depression[J]. Journal of Affective Disorders, 2022, 308(7):141-146. doi: 10.1016/j.jad.2022.04.070
[5]	Li D, Wen G. MRMR-based ensemble pruning for facial expression recognition[J]. Multimedia Tools and Applications, 2018, 77(12):15251-15272. doi: 10.1007/s11042-017-5105-z
[6]	Li D, Wen G, Hou Z, et al. RTCRelief-F:An effective c-lustering and ordering-based ensemble pruning algorithm for facial expression recognition[J]. Knowledge and Information Systems, 2019, 59(1):219-250. doi: 10.1007/s10115-018-1176-z
[7]	Xue Y, Zhu H, Liang J, et al. Adaptive crossover operator based multi-objective binary genetic algorithm for feature selection in classification[J]. Knowledge-Based Systems, 2021(3):107218-107226.
[8]	Dong H, Li T, Ding R, et al. A novel hybrid genetic al-gorithm with granular information for feature selection and optimization[J]. Applied Soft Computing, 2018, 65(4):33-46. doi: 10.1016/j.asoc.2017.12.048
[9]	Zavaschi T, Koerich A L, Oliveira L. Facial expression recognition using ensemble of classifiers[C]. Prague: IEEE International Conference on Acoustics,Speech and Signal Processing, 2011:1489-1492.
[10]	Deb K, Pratap A, Agarwal S, et al. A fast and elitist multiobjective genetic algorithm:NSGAII[J]. IEEE Transactions on Evolutionary Computation, 2002, 6(2):182-197. doi: 10.1109/4235.996017
[11]	张翔, 史志才, 陈良. 基于SWA优化级联网络的表情识别方法[J]. 电子科技, 2020, 33(9):16-20.
	Zhang Xiang, Shi Zhicai, Chen Liang. Expression recognition method based on cascade network optimized by SWA[J]. Electronic Science and Technology, 2020, 33(9):16-20.
[12]	邓源, 施一萍, 江悦莹, 等. 基于MobileNetV2与LBP特征融合的婴幼儿表情识别算法[J]. 电子科技, 2022, 35(8):47-52.
	Deng Yuan, Shi Yiping, Jiang Yueying, et al. Infant expression recognition algorithm based on MobileNetV2 and LBP feature fusion[J]. Electronic Science and Technology, 2022, 35(8):47-52.
[13]	Wang M, Li X, Chen L. An enhance multimodal multi-objective optimization genetic algorithm with special crowding distance for pulmonary hypertension feature selection[J]. Computers in Biology and Medicine, 2022, 146(5):105536-105547. doi: 10.1016/j.compbiomed.2022.105536
[14]	杨东勇, 卢瑾. 基于惩罚机制的自适应交叉粒子群算法[J]. 计算机科学, 2010, 37(4):249-254,284.
	Yang Dongyong, Lu Jin. Self-adaptive crossover particle swarm optimization based on penalty mechanism[J]. Computer Science, 2010, 37(4):249-254,284.
[15]	Goodfellow I J, Erhan D, Carrier P L, et al. Challenges in representation learning:A report on three machine learning contests[J]. Neural Networks:The Official Journal of the International Neural Network Society, 2015, 64(4):59-63. doi: 10.1016/j.neunet.2014.09.005
[16]	Partalas I, Tsoumakas G, Vlahavas I. An ensemble uncertainty aware measure for directed hill climbing ensemble pruning[J]. Machine Learning, 2010, 81(3):257-282. doi: 10.1007/s10994-010-5172-0
[17]	Li N, Yu Y, Zhou Z. Diversity regularized ensemble pruning[C]. Berlin:Machine Learning and Knowledge Discovery in Databases, 2012:330-345.
[18]	Wang Z, Zhao S, Li Z, et al. Ensemble selection with joint spectral clustering and structural sparsity[J]. Pattern Recognition, 2021, 119(11):108061-108074. doi: 10.1016/j.patcog.2021.108061
[19]	Dai Q. A novel ensemble pruning algorithm based on randomized greedy selective strategy and ballot[J]. Neurocomputing, 2013, 122(122):258-265. doi: 10.1016/j.neucom.2013.06.026
[20]	Dai Q, Ye R, Liu Z. Considering diversity and accuracy simultaneously for ensemble pruning[J]. Applied Soft Computing, 2017, 58(9):75-91. doi: 10.1016/j.asoc.2017.04.058
[21]	Tan M, Le Q V. EfficientNet:Rethinking model scaling for convolutional neural networks[C]. Long Beach: Proceedings of the Thirty-sixth International Conference on Machine Learning, 2019:6105-6114.
[22]	Liu Z, Mao H, Wu C, et al. A ConvNet for the 2020s[EB/OL].(2022-01-01) [2022-08-27] https://arxiv.org/abs/2201.03545.

参数	范围
层数量	3~9
卷积核大小	[4 4]~[8 8]
池化核大小	[2 2]~[5 5]
池化步长	2
随机失活率	0.4~1.0
特征映射数量	16,32,48,64,80,96
学习率	0.01,0.03,0.05
维度	7,1 024,2 048,3 072

数据集	FER2013	JAFFE	CK+	RaFD	KDEF
Baseline	0.296 5(231)	0.577 5(231)	0.242 2(231)	0.303 5(231)	0.283 7(231)
UWA-based DHCEP^[16]	0.267 8(53)	0.488 3(134)	0.230 3(142)	0.253 7(5)	0.242 9(113)
DREP^[17]	0.2680(29)	0.493 0(129)	0.230 3(114)	0.265 8(111)	0.242 9(105)
ComEP^[6]	0.267 8(26)	0.488 3(134)	0.230 3(142)	0.265 8(129)	0.248 6(56)
Kappa^[18]	0.287 0(196)	0.530 5(192)	0.241 4(230)	0.279 3(1)	0.283 7(231)
QS^[1]	0.290 6(206)	0.5540(174)	0.241 4(230)	0.279 3(1)	0.283 7(231)
RGSS&B-EP^[19]	0.272 5(118)	0.4930(137)	0.230 9(143)	0.270 0(92)	0.246 9(103)
OO^[5]	0.266 6(49)	0.488 3(134)	0.229 6(142)	0.264 4(129)	0.240 8(53)
SDAcc^[20]	0.267 5(18)	0.488 3(129)	0.230 3(142)	0.248 0(11)	0.242 9(57)
AMGAEP	0.266 0(30)	0.477 1(38)	0.225 0(43)	0.266 5(58)	0.236 3(42)

数据集	KDEF
Baseline	0.274 7(261)
AMGAEP	0.211 3(45)