| Based on facial expression recognition technology,recruiters can use machines to analyze candidates’ facial expressions and evaluate candidates from multiple perspectives.Ultimately,assist the interviewer to make more scientific decisions.Improving the accuracy of facial expression recognition relies on the extraction of stable features.However,the difference between the distribution of the actual environment and the training environment limits the recognition of stable features.Many scholars have used domain adaptation to reduce the distribution difference in training and testing datasets.These studies perform well on visible distribution(i.e.,the distribution of unlabeled test data is visible at training),but most of them deteriorate sharply when the distribution is unknown(i.e.,the distribution of unlabeled test data is invisible at training).In practical applications,facial expression recognition will inevitably be conducted when the distribution is unknown.Therefore,this paper studies the domain generalization problem of facial expression recognition,generalizing the visible distribution to the unknown distribution,and learning a model with better generalization ability to the unknown distribution.The main contributions and innovations of this paper are as follows:(1)A feature difference analysis method for cross-domain dataset of face expressions is proposed.This paper analyzes general public datasets of facial expression recognition.Firstly,statistics and visualization of images and different categories are performed.Then,quantitatively analyze the difference between the datasets by combining qualitative and quantitative analysis.The differences between datasets are analyzed by combining qualitative and quantitative analysis.In detail,the study proposes a method considering visual qualitative analysis to measure the difference of features based on cross-validation of cross-domain datasets.It also designs a method to measure the difference of cross-domain datasets using adversarial validation.The above two methods are compared and cross-validated with the maximum average distance algorithm.The consistency of the results demonstrates their validity,laying the groundwork for followup studies.(2)A domain generalization training strategy for face expression recognition is proposed.To improve the recognition ability of face expression recognition on datasets with unknown distribution,the paper conducts feature difference analysis on datasets with visible distribution.Using datasets with larger feature differences for joint training can effectively extract stable features in cross-domain datasets.This paper then compares it with the joint training dataset consisting of datasets with smaller feature differences.It is found that the proposed training strategy can enhance the domain generalization ability of the facial expression recognition algorithm.(3)A domain generalization model for face expression recognition based on risk extrapolation(REx)is proposed.This paper introduces REx to improve the domain generalization ability of facial expression recognition.REx can extend the known distribution to the unknown distribution and improve the ability of the model to identify robust features.Furthermore,this study improves the loss function for REx.Specifically,the average loss of joint training dataset is replaced by the maximum loss of single dataset.The new loss function enhances the generalization ability to unknown distribution,avoiding the situation that the algorithm over-fits on one dataset and under-fits on other training datasets.Finally,the effectiveness of the algorithm is proved by experiments. |
