基于单一神经网络的实时人脸检测

Real-Time Face Detection Based on Single Neural Network

由于人脸尺度多样性使得人脸检测算法在CPU上运行速度受限,提出了一种新的基于单一神经网络的实时人脸检测算法。首先在网络初始卷积层和池化层中设置较大的卷积核尺寸和步长,缩小输入图像尺寸利于实时检测;然后网络将浅层特征图和深层特征图相融合,增强上下文联系和减少重复检测;最后在多个卷积层上预测人脸位置,利用预测框重叠策略,实现多尺度的人脸检测来提升图像中小尺寸人脸的检测精度。在人脸检测数据集基准和野外标注人脸数据集上测试实验结果表明,本文算法模型精度能够达到92.1%和95.4%。与此同时,本文算法在CPU上实现21帧/s的检测速度。

To improve the limited speed of face detection algorithm on central processing unit (CPU)caused by the diversity of the facescales,we proposed a real-time face detection method based on a single neural network. Firstly, a large convolution kernel and step size were used in the initial convolution and pooling layers, which were able to reduce the size of input images. Then, the shallow and deep feature maps were merged to enhance the context-connection and reduce repeated boxes. Finally, we predicted the face location based on the output of different convolution layers. By using the strategy of overlapping prediction boxes, our method is able to improve the detection accuracy of the smaller size face of input images. Experimental results on face detection dataset and benchmark and annotated face dataset in the wild achieve accuracies of 92% and 95.4%, respectively. Above all, our face detection technique can achieve a high detection speed of 21 frames per second on CPU, which can satisfy real-time detection requirements.