优化的BP网络在面部特征点的深度估计

Depth estimation of facial feature points based on optimized BP neural network

针对在单张面部照片的三维人脸建模过程中因平面照片上各特征点的第三维度数据缺失导致的三维人脸失真的现象,利用BP神经网络预测人脸部各特征点的深度信息.将神经网络用作函数映射,网络的输入为人脸的平面特征点坐标,网络的输出则是特征点对应的深度信息.并采取5种不同优化函数辅助训练BP神经网络的各层权值,同时根据实验结果调整隐含层节点数和网络中各项参数,最终构建一个较为优化的BP神经网络用于特征点深度值的估计.结果表明,经Levenberg-Marquardt(L-M)算法优化后的神经网络估计结果精确度较高,同时网络的稳定性较好.利用优化后的BP神经网络得到的特征点的深度信息可用于在三维人脸建模过程中特征点位置的确定上.

In order to solve the problem that the three-dimensional face distortion is caused by the missing of the third dimension data of the feature points on the single photo, the BP neural network is used to predict the depth information of each feature point of the human face. The neural network is used as the function mapping, the input of the network is the plane coordinate point of the face, and the output is the depth information of the feature point.The weights of each layer of BP neural network are trained by five different optimization functions.The number of hidden layer nodes and the network parameters are adjusted according to experimental re-suits.Finally, a more optimized BP neural network is used to estimate the depth of the feature points.The experimental results show that the neural network estimation result with L M algo rithm is more accurate and the stability of the network is better. Therefore, the depth information of the feature points obtained by the optimized BP neural network can be used to determine the position of the feature points in 3D face modeling.