Research On Facial Expression Recognition Based On Deep Learning

Facial Expression is an important medium for conveying emotional information in the process of human communication,so facial expression recognition has set off an upsurge in the field of computer vision.The existing facial expression recognition technology has been applied in many fields,such as case detection,human-computer interaction,medical diagnosis and recommendation systems.Nevertheless,in some complex scenarios,such as station monitoring,school monitoring,hospital monitoring,etc,the research on facial expression recognition is still a difficult problem.In order to improve the accuracy of facial expression recognition,the focus of the research contents are as follows:First,in order to decrease the sophistication and calculation of the model when practically extracting the facial expression features,a small-scale convolution kernel based on facial expression recognition algorithm is proposed.The algorithm uses small-scale convolution kernels with low computational complexity and fast speed.The entire network structure uses 3 ×3 size convolution kernels to extract features from the input facial expression images,and selects the Softmax function to classify and recognize facial expressions.The experimental results of this article on the public data set CK+show that the algorithm achieves a high accuracy of 97.879%while using fewer parameters.Secondly,due to the limitation of imaging equipment and environmental factors,the collected facial images are low resolution and withdraw facial expression features difficultly,so that the recognition effect of facial expressions is bad.To solve this problem,a low-resolution facial expression recognition system based on super-resolution reconstruction is proposed.The system is composed of two deep neural networks.The first neural network is based on a new hybrid loss function,using a simplified residual network structure superimposed to form a residual block to learn the characteristics of facial images,and reconstruct high-resolution facial expression images with more details;The second neural network uses a small-scale convolution kernel to extract facial expression features of the reconstructed high-resolution image to achieve facial expression classification.On the public data set CK+,the proposed new system was verified and tested.The experimental results show that the system effectively improves the accuracy of expression recognition of low-resolution images of different sizes,including 6×6,12×12,24×24 sizes of reconstructed expression images achieved accuracy rates of 93.838%,96.970%,and 97.374%,respectively.Compared with the unreconstructed expression images,the accuracy was increased by 9.091%,4.647%,and 1.313%respectively.Finally,due to different people interpret the same expression in different ways,it will cause big differences in facial expressions.However,most existing facial expression recognition algorithms cannot well capture discriminative expression features,which is not conducive to improving the accuracy of expression recognition.To solve this problem,an improved facial expression recognition algorithm with channel attention mechanism is proposed.The algorithm uses an improved channel attention mechanism module to achieve the purpose of enhancing important features and suppressing invalid features,thereby further improving the accuracy of the facial expression recognition algorithm.In addition,in order to improve the robustness of the model,the entire model uses the activation function ELU to replace ReLU during the training process.The experimental results of this paper on the public data set CK+illustrate that the algorithm adopts the method of embedding an improved channel attention mechanism module to achieve a facial expression recognition accuracy of 98.384%,which is comparable to the original facial expression recognition without an attention mechanism model.Compared with the algorithm,the accuracy is improved by 0.505%.