Extraction Of Facial Expression Features Under Noisy Scenes And Its Robust Classification Model

The excellent performance of deep learning makes it shine in the field of image recognition,but the advantage is not obvious when it comes to facial expression recognition with low resolution and a lot of interference information.This is due to the external influence of facial expressions such as pose changes and the subjective nature of annotator labeling,which makes it difficult to collect data with accurate labels.In this paper,three facial expression classification models are proposed to address the above mentioned problems.First,to address the problem of poor performance of facial expression recognition algorithms under noisy scenes,a noisy self-healing cleaning facial expression classification model based on a self-attentive mechanism is proposed.The model evaluates the importance weights of the samples by the self-attentive mechanism,and arranges the weights in descending order to highlight the noisy samples,and the relabeling module tries to identify the noisy samples and modify their labels,and finally iteratively updates and retrains the model with purer and purer samples to eliminate the influence of label noise on the deep neural network.Second,to address the problem that traditional convolutional neural networks fail to fully extract the facial information of training samples,a facial expression classification model based on improved octave convolution combined with fast low-rank shared dictionary learning is proposed.The octave convolution algorithm,which can effectively filter the redundant information of the feature map,is optimized to reduce the computation and storage costs while improving the feature extraction ability.Finally,a joint supervised-supervised feature mapping optimization model for robust facial expression classification is proposed to address the problem that supervised learning relies too much on sample labels but ignores information beyond the labels.Self-supervised learning is used to assist supervised learning to avoid the negative impact of noisy labels,and the network parameters are updated by imposing self-supervised learning intrinsic similarity constraint and supervised-self-supervised learning structural similarity constraint on the shared feature encoder respectively to improve the noise immunity performance of the model.