智能设备高昂的设计费用和庞大的计算资源需求成为在便携式、低功耗设备上实现深度学习算法及其应用的主要障碍。文中基于树莓派平台，借助Intel的视频处理器（VPU）低功耗加速模块，设计并实现了基于残差特征提取模块CNN模型的实时人脸检测系统。结果表明，相较于单纯使用树莓派CPU进行计算，文中方法在视频流中检测人脸和人脸关键点提取的实验中分别实现了18.62倍和17.46倍的加速，在便携式设备中实现快速、实时、在线的人脸检测和特征点提取成为现实，同时为使用便携式、低功耗设备运行深度学习算法提供了一种确实可行的方案。

The high design cost and huge computing resource demand of intelligent devices have become the main obstacles to the implementation and application of deep learning algorithm in portable and low-power devices. In order to solve these problems, in this paper, based on the raspberry PI platform and with the help of Intel video processing unit(VPU) low-power acceleration module, a real-time face detection system based on CNN model with residual feature extraction module was designed and implemented. The experimental results show that compared with using central processing unit(CPU) of raspberry PI alone, the proposed method achieved 18.62 times and 17.46 times acceleration respectively in the experiments of face detection and face alignment detection in video stream. It realized the fast, real-time and online face detection and face alignment extraction in portable devices. Meanwhile, it also provided a feasible scheme for the operation of deep learning algorithm in portable and low power devices.

TP391.4

国家重点研发计划资助项目（2018YFB1306603）；国家自然科学基金资助项目（61672436）。