Abstract:

As deep residual network has problems such as complex network structure and high time cost in face recognition applications of small mobile devices， a lightweight model based on deep residual network was proposed. Firstly， by simplifying and optimizing the structure of the deep residual network and combining the knowledge transfer method， a lightweight residual network （student network） was reconstructed from the deep residual network （teacher network）， which reduced the network structural complexity while ensuring accuracy. Then， in the student network， the parameters of the model were reduced by decomposing standard convolution， thereby reducing the time complexity of the feature extraction network. Experimental results show that on four different datasets such as LFW （Labeled Faces in the Wild）， VGG-Face （Visual Geometry Group Face）， AgeDB （Age Database） and CFP-FP （Celebrities in Frontal Profile with Frontal-Profile）， with the recognition accuracy close to the mainstream face recognition methods， the proposed model has the time of reasoning reaches 16 ms every image， and the speed is increased by 10% to 20%. Therefore， the proposed model can have the speed of reasoning effectively improved with the recognition accuracy basically not reduced.

Key words: deep residual network, face recognition, lightweight, Knowledge Distillation (KD), Depthwise Separable Convolution (DSC)

摘要：

针对深度残差网络在小型移动设备的人脸识别应用中存在的网络结构复杂、时间开销大等问题，提出一种基于深度残差网络的轻量级模型。首先对深度残差网络的结构进行精简优化，并结合知识转移方法，从深度残差网络（教师网络）中重构出轻量级残差网络（学生网络），从而在保证精度的同时，降低网络的结构复杂度；然后在学生网络中通过分解标准卷积减少模型的参数，从而降低特征提取网络的时间复杂度。实验结果表明，在LFW、VGG-Face、AgeDB和CFP-FP等4个不同数据集上，所提模型在识别精度接近主流人脸识别方法的同时，单张推理时间达到16 ms，速度提升了10%~20%。可见，所提模型能够在推理速度得到有效提升的同时识别精度基本不下降。

关键词: 深度残差网络, 人脸识别, 轻量级, 知识蒸馏, 深度可分离卷积