基于动态卷积的标签不确定性自学习预测分配算法的面部表情识别-《集美大学学报（自然科学版）》

文章信息/Info

Title:: Facial Expression Recognition Based on Self-Learning Label Prediction and Distribution Algorithm Based on Dynamic Convolutional for Label Uncertainty

作者:: 杨远奇1; 蔡岱立2; 谢泽凌1; 3; 江恩杰1; 1. 集美大学诚毅学院信息工程系，福建厦门 361021；2. 深圳市浩瀚卓越科技有限公司，广东深圳 518071；3. 自然资源部宁德海洋中心, 福建宁德 352100

Author(s):: YANG Yuanqi1; CAI Daili2; XIE Zeling1; 3; JIANG Enjie1; 1.Information Engineering，Chengyi College，Jimei University， Xiamen 361021， China；2.Shenzhen Hohem Technology Co.,Ltd.，Shenzhen 518071，China；3.Ningde Marine Center, MNR,Ningde 352100, China

摘要:: 为解决在表情识别领域中数据集受到噪声、标注模糊、微表情等不确定性因素干扰的问题，提出了一种标签不确定性自学习预测分配算法。该算法包含三个核心模块:1)自注意力加权模块采用动态卷积实现精细的像素级注意力机制,有效降低计算负担；2)正则化排序模块通过样本权重排序和重新分配,优化了模型对不确定样本的处理；3)标签再分配模块对低权重样本进行标签校正,提高了整体预测精度。经过实验验证，该算法有效抑制了标签不确定性的影响，在RAFDB和MMAFEDB等公开数据集上展现了卓越性能。

Abstract:: To solve the problem of uncertain factors such as noise, fuzzy labeling, and micro expressions affecting the dataset in the field of facial expression recognition, a label uncertainty self-learning prediction allocation algorithm is proposed.The algorithm consists of three core modules: 1) A self-attention weighting module employing dynamic convolution to achieve fine-grained pixel-level attention mechanism, effectively reducing computational overhead; 2) A regularized ranking module that optimizes the model’s handling of uncertain samples through sample weight reranking and reallocation; 3) A label reassignment module that corrects labels for low-weight samples, thereby improving overall prediction accuracy. Experimental validation demonstrates the algorithm’s efficacy in mitigating the impact of label uncertainty, exhibiting outstanding performance on publicly available datasets such as RAF-DB and MMAFEDB.Keywords: facial expression recognition; label uncertainty self-learning prediction allocation