Research On Facial Expression Recognition Based On Jetson Nano Platfor

In recent years,with the rise of big data and artificial intelligence,human-computer interaction is becoming a hot research field,and facial expression recognition technology as an important interface for human-computer interaction,it assumes the One-step work for machines understand human emotion through vision,it has important research and application value.Another important research direction of human-computer interaction is how to apply algorithms to more miniaturized embedded devices,it requires solving the contradiction between the limited computing power of embedded devices and algorithms.Face detection technology is the foundation of facial expression recognition technology.This thesis firstly analyzes the face detection algorithm based on Harr-Like features in detail,and explores how to reduce the amount of calculation through the integral graph and through the cascade algorithm,The method of training a strong classifier;then the basic theory and method of convolutional neural networks are introduced in detail,as well as the structure and application scope of several classic convolutional neural networks;and thesis improvements to the classification models of shallow and deep neural networks,make it suitable for facial expression recognition tasks and a comprehensive analysis of their recognition effects.the thesis try to explore a convolutional neural network with higher recognition accuracy and shorter recognition time in facial expression recognition tasks,The model structure was found through experiments,The the 18-layer residual network model recognize the test of 3589 images,it took only 14.29 seconds,and the model only occupies 2.89MB of space,which has well verified that the residual network structure is more suitable for the lightweight improvement of the model.In order to explore a facial expression classification model that can run smoothly on embedded devices,the thesis analyzes the calculation time distribution of convolutional neural networks,and proposes a lightweight residual network based on deeply separable convolution structure Mini-Res model,which requires only 25.45 × 106FLOPs for calculation.It takes 3.16 milliseconds to detect an image on the Jetson Nano device.The space occupied by the model is only 1.68MB.The recognition accuracy rate of the expression recognition test set has reached 58.04%,which basically meets the recognition needs of this article.Finally,this thesis also designs and implements a coupled facial expression recognition system.The system can replace the facial expression recognition model at any time as needed,collect real-time images through the CSI camera,and display the facial expressions in real time on the HDMI display device category;in a video with a resolution of 480 × 480,the fluency can reach up to 30 FPS.