Facial Expression Recognition Based On Deep Attention Network

Facial expression as a signal,that conveys emotion state and intention,plays an indelible role in interpersonal communication,human-computer interaction,safe driving and online education.Recognizing facial expressions efficiently and accurately is a challenging and meaningful task.In recent years,many domestic and foreign scholars and scientific research institutions have conducted in-depth study on it,therefore facial expression recognition has become a hot topic in the field of computer vision.Facial expression recognition refers to the computer classify the extracted facial expression features by the way of human thinking and cognition to analyze and understand human emotion states.The algorithmic of facial expression recognition are generally divided into facial expression recognition based on static pictures and facial expression recognition based on dynamic video sequences.This paper mainly focuses on facial expression recognition in static pictures.Compared with traditional machine learning algorithms which need design and extract features manually,deep learning,a branch of machine learning,is capable of learning multi-level discriminative and robust descriptors in the form of supervised or unsupervised.In recent years,it has achieved remarkable results in computer vision tasks such as object detection,image classification,image retrieval,and semantic segmentation.Various well-performing deep learning models emerge in an endless stream.With excellent performance,deep learning has conquered many researchers and become a popular research method in the field of artificial intelligence.Deep learning technology is utilized to develop related research on facial expression recognition based on static pictures in this paper.After thoroughly researching and continually summarizing the deep learning theory and deep learning model which are well-respected and well-performing,the existing model is improved in this paper.Not only many advantages of the traditional model are inherited,but also the attention mechanism is integrated for strengthening the robustness and accuracy of improved model.The main work of this paper is divided into the following sections:1?In order to minimize the influences caused by the utilization of the pooling layer,a lightweight multi-scale convolutional neural network model is designed in this paper.The model can fuse visual features learned by multiple scale convolution operations effectively.This information can be passed to the learning process of higher layers,different scales features are combined and abstracted simultaneously in next stage.The highly abstract semantic features are learned by hierarchical connections,a characteristic of deep learning,and the accuracy of facial expression recognition is improved by using discriminative and robust semantic features.2?Visual information contained in different regions of the face may have different validity for the facial expression recognition task,while the traditional deep learning based methods ignore this problem and treats the entire face area indiscriminately,and global feature with fixed dimension is extracted from the whole facial expression image.A multi-channel convolutional neural network is proposed in this paper,which is capable of learning global and local features from whole facial images and facial components(eyes,mouth,nose etc.)images respectively.3?In order to alleviate the effect of this problem and redundant noise information on the final expression recognition,this paper introduces the attention mechanism into the deep convolutional neural network to capture different significant areas that make contribution to recognize different facial expressions.The visual information contained in these face regions can be given different weights for effective integrating into discriminative and robust depth features.4?Softmax classification loss is adopted as the supervised signal to train the model optimization parameters in the existing deep learning models,while the softmax classification loss is not capable of satisfying the recognition requirements that increase the intra class similarity and weaken inter class similarity.Hence,two novel loss function,regularized center loss and improved triplet loss,are proposed in this paper which are used in conjunction with softmax respectively as objective function for training models to enhance the intra-class aggregation and inter-class dispersion.At the same time,the discriminability of the features extracted by the deep model is strengthen for improving the accuracy of facial expression recognition.In order to verify the effectiveness of the proposed method,this paper conducts comparative experiments on CK+ and Oulu-CASIA facial expression databases.Experiments show that the proposed methods can significantly improve the accuracy of facial expression recognition.