利用改进深度信念网络的人脸表情识别

Facial expression recognition based on improved deep belief network

针对人脸表情识别过程中,误差逆向传播算法(back propagation,BP)对深度信念网络(deep belief network,DBN)微调时容易陷入极值点局部极小和收敛时间过长的问题,提出基于BP算法微调DBN的改进方法。对表情进行多特征提取并降维,利用所提方法对降维后特征进行学习,采用共轭梯度算法解决BP算法和DBN结合存在的问题。实例仿真计算结果表明,所提方法精度高于支持向量机、基于Wasserstein生成式对抗网和基于图形信号处理的方法。所提方法比卷积类方法络训练时间更短,内存消耗更小。

In the process of facial expression recognition,the back propagation algorithm(BP)is easy to fall into the local minimum of extreme points and the convergence time is too long when the deep belief network(DBN)is fine tuned using BP algorithm.An improved method based on the DBN fine tuned by BP algorithm was proposed.The multi-features of facial expression were extracted and dimensionally reduced.The proposed method was used to learn the reduced feature.The problem of combining BP algorithm with DBN algorithm was solved using conjugate gradient algorithm.Example simulation calculations show that the accuracy of the proposed method is higher than that of the support vector machine method,the method based on Wasserstein generative adversarial network and the method based on graphic signal processing.Compared with the convolutional neural network,the training time of the proposed method is shorter,and the memory consumption is smaller.