摘要
针对监控视频中人脸检测问题,提出了一种基于卷积神经网络的人脸检测方法。利用监控视频中的RGB图像,边缘图像和LBP纹理图像分别作为卷积神经网络的输入,训练三个网络分别提取特征再进行特征融合,最后用SVM训练得到分类器。实验表明,该方法不仅在FDDB数据集上的检测准确率比Viola-Jones检测算子高出了5%,而且在实际监控视中也获得了优于Viola-Jones算子的性能。
As we all know, face detection is an important part of video surveillance, which has been more and more popular over the years. Researchers are eager to put this technique into practical use. However, because of the drop of accuracy, caused by uncontrolla- bility of illumination, pose and expression in realistic scenarios, there is still a sizable gap between the realistic scenarios needs and the current face detection methods. According to the above questions, this paper proposed a new face detection approach, based on convo- lutional neural network. The CNN architecture automatically extracts specific feature from a training set of face and nonfaee patterns. RGB images, edge images and LBP texture images have been used as the input of the CNN to train three different networks respective- ly. From these networks, these specific features are extracted and integrated to a new fusion feature. By training this feature in SVM, this system get a classifier, which can be used to classify the image to be face or nonface. Evaluations on popular face detection bench- mark datasets showed that the proposed approach provides very high detection rate with a particularly low level of false positives . The accuracy rate of the approach is 5% higher than Viola- Jones operator. When it runs in surveillance video, the performance is better than Viola- Jones operator.
出处
《网络新媒体技术》
2016年第6期24-29,共6页
Network New Media Technology
基金
中科院先导课题"海量网络数据流海云协同实时处理系统"(课题编号:XDA060112030)
关键词
CNN
人脸检测
监控视频
CNN, face detection, surveillance video
