Abstract: The convolutional neural network model needs amount of computation to detect faces at different scales. The process consists of several parts, which is too complex. In view of these two deficiencies, a multi-scale convolution neural network is proposed. According to each layer of convolutional neural network with different sizes of receptive field, it extracts the feature vector of a plurality of layers respectively for face classification and regression, and changes the full-connected layers into convolutional layers in order to adapt to different sizes of picture. In this method, multiple steps of face detection are integrated into a convolutional neural network, and the complexity of the model is reduced. Experimental results show that the above improvements can effectively improve the detection performance of the model. Under the same testing conditions, the proposed model has a significant improvement in terms of accuracy and detection speed compared to other face detection models.

Key words: convolutional neural network, face detection, multi-scale, full convolutional network, feature extraction

摘要： 卷积神经网络在检测不同尺度的人脸时所需要的计算量很大，检测过程由多个分离的步骤组成，过于复杂。针对这两方面的不足，提出一种多尺度卷积神经网络模型。根据卷积神经网络各个层具有大小不同的感受野，从不同层提取多个尺度的特征向量分别进行人脸分类与回归，并将网络的全连接层改成卷积层，以适应不同大小的图片输入。该方法将人脸检测的多个步骤集成到一个卷积神经网络中，降低了模型复杂度。实验结果表明，相同测试条件下，所提方法相比其他人脸检测模型在准确率和检测速度上均有显著提升。

关键词: 卷积神经网络, 人脸检测, 多尺度, 全卷积网络, 特征提取