基于改进深度残差网络的低功耗表情识别

Energy-efficient Facial Expression Recognition Based on Improved Deep Residual Networks

为了提高表情识别率并降低表情识别的功耗,提出一种基于改进深度残差网络的表情识别方法。残差学习在解决深度卷积神经网络退化问题、使网络层次大幅加深的同时,进一步增加了网络的功耗。为此,引入具有生物真实性的激活函数来代替已有的整流线性单元(Rectified Linear Units,ReLU)函数,并将其作为卷积层激活函数对深度残差网络进行改进。该方法不仅提高了残差网络的精度,而且训练出的网络权重可直接作为与该深度残差网络具有相同结构的深度脉冲神经网络的权重。将该深度脉冲神经网络部署在类脑硬件上时,其能够以较高的识别率和较低的能耗进行表情识别。

To improve recognition rate and reduce power consumption of facial expression recognition systems,this paper proposed a facial expression recognition method using an improved deep residual networks(ResNets).Residual learning solves the degradation problem of the deep Convolutional Neural Networks(CNNs)to a certain degree and increases the network layers infinitely,but it makes deep CNNs face a more serious power consumption problem.To solve this problem,this paper introduced a new biologically-plausible activation function to improve ResNets and get a facial expression recognition method with both higher performance and lower power consumption.The Rectified Linear Units(ReLU)in the convolutional layers of ResNets are replaced with the new activation function Noisy Softplus.The obtained weights by using the improved ResNets can be directly applied to a deep Spiking Neural Networks(SNNs)architecture derived from the ResNets.The experimental results suggest that the proposed facial expression recognition method is able to achieve higher recognition rate and lower power consumption on a neuromorphic hardware.