孙雪华, 潘晓英. 基于PSO-CG优化深度信念网络的人脸表情识别[J]. 云南大学学报(自然科学版), 2020, 42(5): 877-885. doi: 10.7540/j.ynu.20190471
引用本文: 孙雪华, 潘晓英. 基于PSO-CG优化深度信念网络的人脸表情识别[J]. 云南大学学报(自然科学版), 2020, 42(5): 877-885. doi: 10.7540/j.ynu.20190471
SUN Xue-hua, PAN Xiao-ying. Facial expression recognition using PSO and CG on optimizing the DBN parameters[J]. Journal of Yunnan University: Natural Sciences Edition, 2020, 42(5): 877-885. DOI: 10.7540/j.ynu.20190471
Citation: SUN Xue-hua, PAN Xiao-ying. Facial expression recognition using PSO and CG on optimizing the DBN parameters[J]. Journal of Yunnan University: Natural Sciences Edition, 2020, 42(5): 877-885. DOI: 10.7540/j.ynu.20190471

基于PSO-CG优化深度信念网络的人脸表情识别

Facial expression recognition using PSO and CG on optimizing the DBN parameters

  • 摘要: 深度信念网络(Deep Belief Network,DBN)在人脸表情识别领域表现出较好的性能,但其在有监督学习阶段通常采用反向传播微调网络的初始参数空间,容易陷入局部最优值,不能找到网络的最优参数空间. 提出一种粒子群算法(Particle Swarm Optimization,PSO)和共轭梯度(Conjugate Gradient,CG)融合进行优化DBN,将其用于人脸表情识别. 首先,通过虚拟样本扩充训练样本数据集的方式构建了人脸表情数据集. 然后,使用PSO寻找全局最优点,DBN的参数维度高,单独使用PSO优化网络参数空间,容易出现局部最优、收敛速度慢等问题, 结果显示,PSO与CG相融可以加强搜索精度和加快收敛速度,获取更优模型参数. 仿真结果表明,提出的方法在JAFFE和CK+这2个常用人脸表情数据库的识别率分别为94.52%和97.84%.

     

    Abstract: Deep Belief Network (DBN) has good performance on facial expressions recognition, but it usually utilizes back propagation to refine the initial parameters of the network in the supervised learning stage. Unfortunately, it is easy to fall into the local optimal only with back propagation approach, and can not find an optimal network parameters for deep belief network. In order to address this problem, a fusion of Particle Swarm Optimization (PSO) and Conjugate Gradient (CG) is proposed to optimize DBN in the supervised learning stage. Firstly, the virtual sample generation method is used to expand training sample data set. Then, the fusion algorithm utilizes PSO to seek the global optimum of the search space. But the high dimension of DBN, using PSO alone to optimize the network parameter space is prone to the local optimization, slow convergence and the other problem.Finally, integrating with CG can enhance the PSO's convergence precision and accelerate its convergence rate,and model parameters may obtain better training. The results show that the proposed method achieves the recognition rates of 94.52% on the JAFFE database and 97.84% on the CK+ database.

     

/

返回文章
返回