Research On Facial Expression Recognition Method In The Wild

As one of the important research directions in the field of affective computing,facial expression recognition is widely used in human-computer interaction systems such as robot interaction,medical diagnosis and safe driving.The current facial expression recognition varies greatly due to different scenes.The facial expressions in the lab are mostly frontal,clear,and performative images.The related algorithms are relatively mature and the recognition accuracy is high.Since the facial expressions in the wild are closer to the actual application of the project,their facial expressions have the characteristics of data imbalance,occlusion,and pose variations.These characteristics make the application of facial expression recognition still challenging.At present,there are few studies on the imbalance of facial expression data,and the recognition accuracy of related facial expression recognition algorithms on occlusion and pose variations datasets is also low.Therefore,this paper mainly studies the data imbalance,occlusion and pose variations problems of facial expression recognition in the wild.The main work is as follows:(1)Aiming at the problem of data imbalance in facial expression recognition in the wild,a data augmentation method based on generative adversarial networks and active learning is proposed.This method uses Star GAN to generate facial expression samples.For the noise problems in the generated samples,active learning is used to filter the more informative samples,and the data expression of a small number of expression categories is augmentated.At the same time,the channel attention module combines with the residual network to serve as a base classifier for facial expression recognition.The data augmentation method proposed in this paper is experimentally verified on the RAF-DB dataset,and the results show that its recognition accuracy reaches 85.16%,which has a better perfermance than other data imbalance solutions.(2)Aiming at the problems of occlusion and pose variations in the facial expression recognition in the wild,an expression recognition model based on local feature enhancement with weakly supervised learning is proposed.The model uses weakly supervised learning to obtain attention feature maps,and enhances local features through attention feature maps;since second-order features can capture the degree of facial distortion,covariance pooling is added to the bilinear attention module to extract second-order statistics of faces quantitative features;and adopts metric loss Arc Face Loss as the target loss function of the model to solve the problem of small differences between classes and large differences within the class of facial expressions,improving the model's ability to learn discriminative features.Through experimental verification,it is shown that the expression recognition models proposed in this paper achieve recognition accuracy of 87.04%,61.05%,and 67.39% on RAF-DB,Affect Net,and FED-RO,respectively,which are better than the experimental results of the comparative literature.And through cross-dataset experiments,the model is verified to have good generalization performance.In summary,the data augmentation method proposed in this paper can effectively solve the problem of data imbalance in facial expression recognition in the wild,thereby improving the accuracy of facial expression recognition;the facial expression recognition model based on local feature enhancement with weakly supervised learning proposed in this paper achieve better recognition results in facial expression datasets in various scenes.