基于神经网络集成的多视角人脸识别

人脸在图像深度方向上发生偏转时，即使同一对象的人脸图像也会发生极大的变化。在此，将神经网络集成应用于多视角人脸识别，所用的人脸特征通过多视角特征脸分析获得。为每一视角的特征空间各训练一个神经网络，并利用另一个神经网络对其进行结合。利用训练好的神经网络集成进行识别时不仅不需进行偏转角度估计预处理，而且还可以在给出识别结果的同时给出角度估计信息。实验结果表明，该方法的识别精度高于根据精确的偏转角度估计信息挑选最佳单一神经网络所能达到的效果。

VIEW-INVARIANT FACE RECOGNITION BASED ON NEURAL NETWORK ENSEMBLE

When human faces rotate in image depth, even the faces of the same person appear with great variances. In this paper, neural network ensemble is applied to view invariant face recognition. The facial features used are extracted through view specific eigenface analysis. Several neural networks are trained, each for an eigenspace of different views, and their results are combined with another neural network. After the ensemble is trained, view estimation is not required for recognition. Moreover, when new faces are fed, the ensemble will not only give the recognition result but also present an estimated view information. Experimental results show that the recognition accuracy of the proposed approach is better than that of the best individual neural network selected according to the information provided by an accurate front end view estimation process.