基于改进BP网络的人脸检测与定位

Human Face Detection and Location Method Based on Improved BP Network

提出了一种基于改进BP网络进行人脸检测与定位的方法,采用变步长的学习速率,在加快学习速度的同时,保证了权值的稳定性;采用加动量项的BP算法,减小了权值的振荡,且可以避免网络陷入局部最小。利用图像的灰度信息对已构建好的神经网络进行训练,然后利用已训练好的神经网络进行搜索,确定被检测的窗口是否包含人脸。实验结果表明此方法比传统的人脸检测与定位方法具有更强的鲁棒性和可扩展性,定位速度快,泛化能力显著。

A human face detection and location method based on improved back-propagation network is proposed. Modificative learning rate is introduced in order to increase the learning speed and ensure the weights＇ stabilization. BP algorithm accessioned momentum is adopted to minish the weights＇ surge and avoid the local least of network. The built neural network is trained by use of image＇ grey information and BP algorithm accessioned momentum, then the neural network which has been trained are used to search an image and decides whether detected windows of the image contains a face. The experimental results show this method possesses higher robustness and expansion than traditional human face detection and location method, and the speed of location is very quickly and evolvement ability is rather marked.