Cross-database Expression Recognition With Feature Completeness Learning

With the flourishing progress of artificial intelligence technology,recent expression recognition systems have achieved impressive success.The promising performance of existing algorithms depends on the distribution consistency between the training dataset and the testing dataset.However,many factors such as collecting conditions and identity characteristics,lead to various data.Hence most expression recognition methods suffer from accuracy degradation when being allied to a new dataset with different feature distribution.Therefore,cross-dataset expression recognition has important academic value and the solution to this problem is beneficial to the widespread application of expression recognition.Most of the existing cross-dataset expression recognition researches lack the effective elimination of interference information,leading to domain information and neutral information in extracted features.The lack of completeness of facial features is the main reason for the degradation of performance.Therefore,how to extract more complete and robust expression features is the key to tackle cross-database expression recognition.Based on deep convolution neural network,this thesis aims to build a cross-database expression recognition system and focuses on learning a feature space with more complete expression features and fewer interference features.We explored from the perspective of reducing domain features and neutral features,which effectively improves the recognition performance.The main work of this article can be described as follows:1)Cross-database facial expression recognition with domain-invariant and discriminative feature learningThis thesis improves the completeness of expression features by reducing domain information.The maximum mean discrepancy is calculated to measure the distribution distance,which assists to learn features that are robust to domain variations.Simultaneously,cosine similarity is conducted to guarantee the discriminative ability of expression features.The combination of distribution consistency constraint and feature distance constraint allows the model to concentrate on expression information while removing interference information,which facilitates the learning of complete features.The effectiveness of this method is validated through experiments.2)Residual feature learning for cross-database expression recognitionThis thesis improves the completeness of expression features by reducing neutral information.In the proposed residual feature learning framework,the facial expression information extraction branch and the neutral information extraction branch share low-level convolution layers.The residual features are obtained by calculating the difference between the expression information and neutral information.Residual features are more complete and robust for the expression recognition task.The experiments validate the superiority of our method over other methods.3)Propose the cross-database expression recognition framework based on domain alignent and residual featre learningThe domain alignment method and the residual feature learning method are described from different perspectives to improve the completeness of expression features.The domain alignment method is considered from the overall domain distribution,and the residual feature learning method is considered from individual features.Therefore,these two methods are unified into one framework to reduce domain information and neutral facial information.This combination makes the model have sufficient representations for expression information and better resistance to interference information.The performance is proved by the experiments.