Research On Semi-Supervised Expression Recognition Algorithm Based On Contrastive Learning

In real human life,facial expressions play an irreplaceable role in intuitive emotional interaction.Although more and more research work relies on supervised learning of artificially labeled samples to realize expression recognition,in reality,human facial expressions are complex and changeable,and emotional features are easily confused.Even for humans,it is difficult to accurately distinguish complex facial expressions.It is also a challenging task,which leads to the high cost of manual annotation,and the shortage of high-quality labeled samples.At the same time,contrastive learning has attracted the attention of many researchers because of its excellent feature discrimination ability and excellent performance in unsupervised classification tasks.Therefore,how to use contrastive learning to improve the discriminative performance of the expression recognition model,and how to introduce deep semi-supervised learning to reduce the dependence on labeled samples,while focusing on samples without human participation to train a high-performance semi-supervised expression recognition model has great significance in the current research.In this paper,the facial expression recognition algorithm based on contrastive learning and semi-supervised learning is studied,and the following work is carried out:(1)Aimimg at the significant intra-class differences and inter-class similarities of realistic expressions,as well as the issue of category imbalance,this paper proposes an Extra-Contrast Affinity Network(ECAN)algorithm based on supervised contrastive learning.ECAN consists of a feature processing network and two newly proposed deep metric loss functions.The feature processing network provides current and historical embedding features to provide the necessary conditions for the algorithm to run.Extra Negatives Supervised Contrastive loss(ENSC loss)considers multiple positive examples and a large number of extra negative examples at the same time to maximize the intra-class compactness and inter-class separation of embedding features,while automatically adjusting how much the model pays attention to majority classes and minority classes to implicitly alleviate the class imbalance issue.Multi-View Affinity loss(MVA loss)improves the center loss with historical feature groups to dynamically learn more accurate class centers and further enhance intra-class compactness.Experimental results on widely used realistic expression datasets such as RAF-DB verify the excellent discriminative performance and effectiveness of ECAN.(2)Aiming at the confidence threshold setting issue and the utilization efficiency of unlabeled samples generated by using the existing semi-supervised learning algorithm for facial expression recognition,this paper proposes a Weighted Adaptive Pseudo-Negatives(WAPN)expression recognition algorithm based on semi-supervised contrastive learning.The WAPN algorithm proposes a weighted adaptive confidence threshold to automatically measure the training difficulty of various expressions and set an adaptively adjusted dynamic threshold to generate pseudo-labels,and then combine the consistency regularization for pseudo-label learning.In addition,a Pseudo Negatives Contrastive loss(PNC loss)is also proposed.This loss combines pseudo-labels to achieve semi-supervised contrastive learning,which alleviates the negative impact of false negatives and effectively improves the utilization efficiency of unlabeled samples.The experimental results on commonly used expression recognition datasets prove the semi-supervised learning ability of the WAPN algorithm.(3)Combining the above achievements,a facial expression recognition system was developed to evaluate the performance difference of realistic expression recognition.This system combines the fully-supervised/semi-supervised recognition model trained by the improved MTCNN and ECAN/WAPN algorithm to realize the system flow from realistic expression to emotion classification.Through static and robustness system comparison experiments,the advantages and differences of the fully supervised system and the semi-supervised system in recognizing realistic expressions are verified and analyzed,which proves that the research results of this paper have certain practical application value.