基于选择性搜索和卷积神经网络的人脸检测

针对复杂背景下存在的光照变化及多姿态的人脸检测问题,提出一种基于Gabor优化的卷积神经网络和选择性搜索策略相结合的算法进行人脸检测。首先采用选择性搜索策略检测出图像中可能存在人脸的目标候选窗口,然后,将候选窗口中的图像子块作为训练好的改进的卷积神经网络的输入,经过一系列卷积和池化操作后,提取窗口图像的特征信息并进行分类,确认候选窗口中是否包含人脸。算法在LFW人脸数据库上取得较高的检测率及检测速度,实验结果表明融合Gabor特征的卷积神经网络用于人脸检测时可避免传统手工提取特征造成的不确定性,具有更好的泛化能力及鲁棒性。

Face detection based on selective search and Gabor optimizing convolutional neural network

To solve the problem of detecting faces with large variances on pose and illumination under complex background, this paper presented a robust and fast algorithm which combined selective search strategy with Gabor optimizing convolutional neural network. Selective search strategy selected the candidate regions. Image warping computed a fixed-size convolutional neural network input from each region proposal. After a series of convolution and pooling operations, convolutional neural network extracted features of candidate regions to confirm whether candidate regions contained faces or not. Experiments on LFW database show that Gabor optimizing convolutional neural network can avoid the uncertainty of traditional feature extraction and have better generalization and robustness.