Abstract: Biometric identification systems should have fast and accurate classification capabilities. Aiming at the problems of traditional face detection methods, such as single feature extraction and long training time, gradient easy to disappear and over-fitting based on deep learning algorithm, a novel face detection algorithm BM-CNN(based on mixnetwork-convolutional neural network) is proposed. The algorithm firstly uses human face segmentation and image enhancement based on curvature filtering to preprocess human face image, and then uses the optimized convolutional neural network (CNN) to preprocess image feature extraction and decision classification. For the convolutional neural network, a new parallel convolutional network and a new pooling strategy are proposed. CNN uses double-mean pooling strategy and a batch normalization (BN) method and multiple types of nolinear units to improve the algorithm detection performance. BM-CNN detects the human face through the double-line convolutional neural network strategy. Finally, this paper conducts comparative experiments on NUAA and CASIA datasets. The experimental results show that the algorithm can classify the face images accurately and also has some improvement in terms of sample size and training time.

Key words: biometric feature recognition, curvature filter, double-line convolutional neural network (CNN), double-mean pooling, batch normalization

摘要： 生物特征识别系统必须拥有快速准确的分类能力。针对传统人脸活体检测方法的特征提取单一和基于深度学习的检测算法中的网络训练时间长、梯度容易消失以及过拟合等问题，提出一种新型人脸活体检测算法BM-CNN（based on mixnetwork-convolutional neural network）。算法首先采用人脸分割技术和基于曲率滤波的图像增强技术对人脸图像进行预处理，然后使用优化卷积神经网络（convolutional neural network，CNN）对预处理图像进行特征提取与决策分类。对卷积神经网络，提出一种复合的并行卷积神经网络，CNN使用二均值池化策略，并综合批量归一化BN（batch normalization）方法和多类型非线性单元提高算法检测性能，通过双线并行的卷积神经网络对活体人脸进行检测。在NUAA数据库和CASIA数据库上对算法进行对比实验，实验结果显示该算法能对人脸图像进行准确的分类，并在样本数量和训练时间上有较大的提升。

关键词: 生物特征识别, 曲率滤波, 并行卷积神经网络, 二均值池化, 批量归一化