基于深度学习的人脸识别算法研究

发布时间：2018-05-29 23:42

本文选题：深度学习 + BP神经网络　；参考：《兰州交通大学》2017年硕士论文

【摘要】：人脸识别是生物识别的一个重要研究方向,随着众多学者的不断努力和长期探索,人脸识别取得了诸多成就,但对于问题的彻底解决还存在一定的距离。近年来诸多学者在神经网络的基础上提出了深度学习,由于其超强的学习能力,现阶段已成为神经网络的主要研究方向。深度学习的提出,为人脸识别问题的彻底解决提供了新的思路。本文利用深度学习算法进行人脸识别,进一步改善人脸识别效果。本文的主要研究内容为:(1)本文首先利用现阶段性价比最高的BP神经网络(Back Propagation,BP)来进行人脸识别研究。针对BP网络在人脸识别过程中,人脸图像数据量大和易陷入局部最优的问题,提出利用主成分分析法(Principal Component Analysis,PCA)和遗传算法(Genetic Algorithm,GA)来优化BP网络,构成PCA-GA-BP网络。该网络首先利用PCA算法来处理人脸图像,减少人脸图像数据量,再通过GA算法对BP网络进行优化,提高网络性能,最后利用AR数据库和ORL数据库进行实验。实验结果表明,该算法不仅可以克服BP网络的缺陷,还能进一步提高人脸识别精度。(2)针对PCA-GA-BP网络在训练样本逐渐增大的情况下,其学习能力小范围下降的问题,提出利用具有超强学习能力的深度信念网络(Deep Belief Networks,DBNs)来替换BP网络,构成PCA-GA-DBNs网络。该网络首先利用GA算法和Gibbs采样来实现网络的逐层训练,训练完成后再利用BP网络进行微调,使其成为最优网络。然后通过AR数据库和ORL数据库进行实验,实验结果表明,PCA-GA-DBNs网络能够很好的提高人脸识别精度,实验最后还分析了不同分类器对人脸识别结果的影响。(3)针对在较大训练样本情况下,GA算法爬山能力不足,容易出现早熟收敛的问题,提出利用全局搜索能力更强且不会陷入局部最优的模拟退火遗传算法(Simulated Annealing Genetic Algorithm,SAGA)来代替GA算法,构成PCA-SAGA-DBNs网络。该网络利用SAGA算法结合Gibbs采样来逐层训练网络,训练完成后再利用BP网络对其进行微调并构造分类器。然后以AR数据库和ORL数据库为实验对象,实验结果表明该网络可以很好的克服GA算法爬山能力不足早熟收敛的缺陷,提高人脸识别精度。最后将本文改进的三种算法进行实验,通过实验结果比较得出PCA-SAGA-DBNs网络不仅具有良好的识别效果,还具有较好的稳定性,是一种较优的人脸识别方法。
[Abstract]:Face recognition is an important research direction of biometrics. With the continuous efforts and long-term exploration of many scholars, face recognition has made many achievements, but there is still a certain distance to solve the problem thoroughly. In recent years, many scholars have put forward deep learning on the basis of neural network. The stage has become the main research direction of neural network. Advanced learning provides a new idea for the complete solution of face recognition. This paper uses depth learning algorithm to carry out face recognition to further improve the effect of face recognition. The main contents of this paper are as follows: (1) this paper first uses the BP nerve with the highest cost performance at the present stage. Back Propagation (BP) for face recognition research. Aiming at the problem of large amount of face image data and easy to fall into local optimal in the process of face recognition in BP network, the principle component analysis (Principal Component Analysis, PCA) and genetic algorithm (Genetic Algorithm, GA) are used to optimize the BP network. Collaterals first use PCA algorithm to deal with face images, reduce the amount of face image data, and then optimize the BP network by GA algorithm to improve the network performance. Finally, the experiment is carried out using the AR database and ORL database. The experimental results show that the algorithm can not only overcome the lack of BP network, but can further improve the accuracy of face recognition. (2) P In the case of the gradual increase of training samples, CA-GA-BP network has the problem of decreasing learning ability, and proposes to replace BP network by using the Deep Belief Networks (DBNs) with super learning ability to form a PCA-GA-DBNs network. The network first uses GA algorithm and Gibbs sampling to train the network by layer by layer training and training. After the completion of the practice, the BP network is used to make the fine-tuning to make it the best network. Then the experiment is carried out through the AR database and the ORL database. The experimental results show that the PCA-GA-DBNs network can improve the face recognition accuracy well. Finally, the experiment also analyzes the influence of different classifiers on the face recognition results. (3) for the larger training sample situation, the experiment results are also analyzed. Under the condition of the GA algorithm, the ability to climb mountains is insufficient and the problem of premature convergence is easy to appear. It is proposed that the simulated annealing genetic algorithm (Simulated Annealing Genetic Algorithm, SAGA), which makes use of the global search ability and will not fall into the local optimal, to replace the GA algorithm to form the PCA-SAGA-DBNs network. The network uses SAGA algorithm to combine Gibbs sampling to layer by layer. Training network, after the completion of training, the BP network is used to fine tune it and construct the classifier. Then the AR database and the ORL database are used as the experimental objects. The experimental results show that the network can overcome the defects of the premature convergence of the climbing ability of the GA algorithm and improve the accuracy of face recognition. Finally, the three improved algorithms in this paper are implemented. It is proved that the PCA-SAGA-DBNs network not only has good recognition effect, but also has good stability. It is a better face recognition method.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】