基于Xception的实时情绪识别在课堂质量分析上的研究

Research on Attention Perception of Real-time Emotion Recognition Based on Xception Architecture

长久以来,因梯度弥散与参数规模较大,深层神经网络在轻量级嵌入式设备上的应用并不多。本文提出了一种基于深度可分离卷积思想的Xception模型,在轻量级嵌入式平台实现了高性能的实时视频识别,并首次通过对泛化的人脸和表情分析感知人们的注意力信息,应用于课堂教学质量分析。在ARM平台上相比普通CNN模型大幅度缩小了CPU、GPU等资源占用,证明实时脸部表情评估可在如nVIDIAJetsonNano等小型计算终端上应用于课堂教学质量分析等场景。

Due to the large amount of parameters and gradient dispersion,the application of DNN on small-scale terminals is uncommon for a long time.This paper proposes a specific application of the Xception model based on the depth separable convolution idea for real-time video recognition in a lightweight terminal,and for the first time,by using the generalized face and expression analysis to perceive people's attention information,it is applied to thefield of teaching quality analysis in classroom.Then it analyzes and verifies the model on the ARM platform.Compared with the ordinary CNN model,the CPU and GPU resources occupied are greatly reduced,consequentlyreal-time facial expression evaluation becomes possible on small computing terminals such as the NVIDIA Jetson Nano.