基于异构多核并行加速的嵌入式神经网络人脸识别方法

Embedded Neural Network Face Recognition Method Based on Heterogeneous Multicore Parallel Acceleration

针对传统视频监控设备进行前端人脸识别时处理大量人脸数据所面临的计算性能不足的问题,提出了一种基于CPU-多核加速器异构结构的前馈神经网络并行加速框架,然后借助主成分分析方法对人脸数据进行特征提取用于神经网络的训练,并将训练好的神经网络模型导入神经网络加速框架中进行分类识别的方法。该方法最终在集成Zynq SoC和Epiphany的Parallella嵌入式并行计算平台中进行了系统实现。实验数据表明,该方法在保证识别准确率一致的情况下,能够提供相对于Zynq中的双核ARM处理器8倍的识别加速能力,在嵌入式人脸识别加速方面具有显著作用。

Computing performance for massive face data is one of the key problems for face recognition on surveillance device.To improve the performance of embedded face recognition systems,a novel parallel feed forward neural network acceleration framework was established based on CPU-multicore accelerator heterogeneous architecture firstly.Secondly,a feature extraction method based on PCA algorithm was used to extract face features for neural network training and classification.Thirdly,the trained neural network parameters can be imported to the parallel neural network framework for face recognition.Finally,the architecture was implemented on hardware platform named Parallella integrating Zynq Soc and Epiphany.The experimental results show that the proposed implementation obtains very consistent accuracy than that of the dual-core ARM,and achieves 8 times speedup than that of the dual-core ARM.Experiment results prove that the proposed system has significant advantages on computing performance.