| Facial expression is an important tool for humans to achieve efficient communication and interaction.Research on facial expression recognition is of great significance.However,label ambiguity and label noise are commonly present in real-world expression datasets.In recent years,label distribution learning(LDL)has emerged as a solution to alleviate label ambiguity.However,some existing LDL-based methods often use Gaussian functions with fixed variances to construct label distribution(which is called fixed-form label distribution),which limits the representational capacity of label distribution because same-class instances share the same label distribution.Additionally,the reliability of constructed label distributions is often ignored in existing methods,which may further introduce label noise and affect model performance.To address these issues,this thesis investigates LDL-based facial expression recognition and makes the following contributions:1.Innovatively proposed a novel adaptive label distribution learning method based on disentangled expression embedding and established a sequential representation of facial expression labels.Based on this,a strategy for adaptively constructing label distribution for expression instances is proposed to address the difficulty of existing fixed-form label distribution learning methods to adequately represent the difference between similar instances in expression recognition task.In addition,a trainable two-branch network is designed to separate the face identity attributes from the source face images to ensure robust facial expression embedding is obtained.2.Innovatively proposed a self-paced label distribution learning method to address the issues of label ambiguity and label noise in facial expression recognition.This method adopts a simple yet efficient label distribution generator to solve the problem of label ambiguity and uses the generated label distribution as the ground-truth label distribution to guide label distribution learning.Meanwhile,to alleviate the problem of label noise,a self-paced learning paradigm is introduced to mimic the human learning process by gradually increasing the difficulty of training data to enhance the model’s performance.Thus,a self-paced label distribution learning strategy is proposed to improve the model’s robustness and generalization ability by selecting training data reasonably.3.Innovatively proposed a robust label distribution learning method for facial expression recognition.This method is based on the assumption that facial images and their neighboring images have similar feature distributions in the feature space,and constructs a reasonable KNN-Graph to represent adjacency information.Then,a neighbor-guided training loss is proposed,which can effectively alleviate the issues of label ambiguity and label noise in facial expression recognition task,thereby improving the model’s robustness and generalization ability.The experimental results on real-world facial expression datasets such as RAF-DB and Affect Net show that the facial expression recognition methods based on label distribution learning studied in this thesis have significant effectiveness and rationality.The three proposed methods can effectively solve the problems of label ambiguity and label noise in facial expression recognition,and are applicable to facial expression recognition in real-world scenarios. |
