| Compared to ordinary facial expressions,micro-expressions are momentary and subtle changes in facial expressions that last for a very short period,typically less than a second.They are often imperceptible to the naked eye and require high-frame-rate cameras to record.Micro-expressions are usually produced involuntarily and are difficult to suppress or conceal,making them more reflective of genuine emotions.Therefore,analyzing micro-expressions can be a valuable technique for determining an individual’s internal states and emotional changes.Since facial expression features almost exclusively appear in specific regions of the face,using the entire face as the model’s input would result in the model learning identity information unrelated to micro-expression features.This not only increases the feature learning task but also reduces the model’s overall recognition ability.Moreover,due to the limited sample size of existing micro-expression datasets,directly applying deep learning methods to micro-expression recognition can lead to overfitting.Additionally,microexpressions have the characteristic of small facial movements,making it difficult for networks to extract clear micro-expression features from micro-expression samples,resulting in low recognition accuracy.To address these issues,this paper proposes a micro-expression recognition method based on feature disentanglement.The main research work is as follows:(1)A feature disentanglement based micro-expression recognition method is proposed to address the issue of interference caused by identity information that is not related to facial expression features,which leads to inaccurate micro-expression recognition.The method first preprocesses micro-expression samples by face cropping and Eulerian video magnification algorithm.Then,a feature disentanglement network model based on adversarial learning is constructed,which not only eliminates the interference of identity features that are irrelevant to micro-expression recognition,enabling the model to focus more on the crucial micro-expression features,but also separates the different features between multi-task networks to improve the model’s feature extraction ability and robustness.Finally,the performance of the model is verified using test samples from the microexpression datasets SMIC and CASME II.The experimental results show that the recognition accuracy of the model on the SMIC and CASME II datasets reaches 65.91% and 69.31%,respectively,which is 4.62 and 4.49 percentage points higher than the baseline model.(2)Micro-expressions have a short duration,small facial muscle movement amplitude,and insufficient sample size in existing micro-expression datasets.Applying deep learning methods directly to automatic micro-expression recognition can lead to issues such as model overfitting.To address these issues,transfer learning is introduced to assist micro-expression recognition by utilizing rich and well-annotated general facial expression data.By leveraging knowledge from the source model to enhance learning in the target task,the accuracy of micro-expression recognition can be improved.Additionally,different facial expression images of individuals are used as input to the model to expand the number of micro-expression samples and partially solve the problem of insufficient sample size in micro-expression datasets.The experimental results show that after using this method,the feature disentanglement network model achieves recognition accuracies of 68.33%and 71.20% on the SMIC and CASME II micro-expression datasets,respectively,which is an improvement of 2.42 and 1.89 percentage points compared to the method without transfer learning.(3)To address the problem of the small facial muscle movement amplitude in microexpressions,which makes it difficult for the facial expression feature extractor to extract clear microexpression features,a micro-expression recognition model based on feature disentanglement and adversarial learning is proposed.Since micro-expression images usually have ordinary expression images of the same category,the adversarial learning method is used to effectively learn the shared features between micro-expressions and ordinary expressions,making similar features closer to each other and different features farther away from each other.Then,the loss inequality regularization is used to correct the output space expression classification loss of the ordinary expression-identity feature disentanglement network model and the micro-expression-identity feature disentanglement network model.This method can effectively improve the representation learning ability of the feature disentanglement and adversarial learning-based micro-expression recognition model and the accuracy of micro-expression recognition.The experiment shows that the micro-expression recognition model based on feature disentanglement and adversarial learning achieves an accuracy of72.67% and 75.80% on the SMIC and CASME II datasets,respectively. |
