Research On Facial Expression Recognition Based On Deep Neural Network

With the rapid development of artificial intelligence and machine vision,facial expression recognition,as an important part of human computer interaction,has a great application prospect.However,most of the methods use 2D static images or 2D video sequences combined with traditional classifiers to complete the recognition task.The limitation of this method is that the classification accuracy often relies too heavily on classifier's parameters and 2D images are sensitive to pose and illumination.To counter the problems above,we present two convolutional neural network(CNN)-based facial expression recognition system.This paper mainly classified static facial expression images.The main research contents are as follows:We construct a single-channel double-input CNN framework to prove that the classification ability of neural network is better than traditional classifiers and introduce local binary pattern(LBP)to capture facial texture information to improve the network performance.The current representative CK+,JAFFE and FER2013 facial expression datasets are used to verify the effectiveness of the framework.The recognition accuracy was 97.34%,98.92% and 67.71%,respectively.Meanwhile,three parallel experiments were designed to compare the performance of the proposed framework.Aiming at the limitations of 2D images,we constructed a dual-channel double-input CNN framework,and collected a new RGB-D(color + depth)facial expression dataset using Kinect sensor.The dataset included seven expressions(neutral,angry,disgust,fear,happy,sad and surprise)from 15 subjects(9 male and 6 female)aged from 20 to 25.The collected RGB-D images are preprocessed by image normalization,pixel filtering,grayscale rendering and facial extraction to maximize the use of the images.The experimental results demonstrate that the proposed network with RGB-D images are superior to the one with only RGB images or depth images.