Scholarlink

<p>Facial Expression Recognition (FER) refers to the automated analysis of human emotional states using computer vision techniques, which are of significant importance in tasks such as fatigued driving detection, learning engagement analysis, and safety monitoring. In recent years, deep learning methods have made remarkable progress in facial expression recognition tasks. However, due to scarce samples, high intra-class variation, and inter-class similarity, current approaches perform poorly in recognizing difficult categories, such as Disgust and Fear. To alleviate the performance degradation resulting from these difficult categories, we propose a training framework called Contrastive Learning Distillation of Non-target Categories (CLDN). The proposed method consists of two stages: In the first stage, the model is jointly trained using contrastive learning and supervised learning to weaken the influence of imbalanced labels on visual representation learning, thereby generating robust soft labels. In the second stage, simple category knowledge is transferred to the difficult categories through self-distillation. During the distillation phase, a Non-target Categories Distillation Loss is introduced, alleviating the inhibition of classic distillation losses on non-target category knowledge and further promoting knowledge transfer during the distillation process. The proposed approach achieves competitive results on the RAF-DB, FERPlus, and AffectNet datasets. Ablation studies demonstrate that the method significantly improves the network’s recognition accuracy for difficult categories. The code is available at <a href="https://github.com/Greysahy/CLDN">https://github.com/Greysahy/CLDN</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Contrastive learning distillation of non-target categories for facial expression recognition

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Heng-Yu An,
  • Rui-Sheng Jia

摘要

Facial Expression Recognition (FER) refers to the automated analysis of human emotional states using computer vision techniques, which are of significant importance in tasks such as fatigued driving detection, learning engagement analysis, and safety monitoring. In recent years, deep learning methods have made remarkable progress in facial expression recognition tasks. However, due to scarce samples, high intra-class variation, and inter-class similarity, current approaches perform poorly in recognizing difficult categories, such as Disgust and Fear. To alleviate the performance degradation resulting from these difficult categories, we propose a training framework called Contrastive Learning Distillation of Non-target Categories (CLDN). The proposed method consists of two stages: In the first stage, the model is jointly trained using contrastive learning and supervised learning to weaken the influence of imbalanced labels on visual representation learning, thereby generating robust soft labels. In the second stage, simple category knowledge is transferred to the difficult categories through self-distillation. During the distillation phase, a Non-target Categories Distillation Loss is introduced, alleviating the inhibition of classic distillation losses on non-target category knowledge and further promoting knowledge transfer during the distillation process. The proposed approach achieves competitive results on the RAF-DB, FERPlus, and AffectNet datasets. Ablation studies demonstrate that the method significantly improves the network’s recognition accuracy for difficult categories. The code is available at https://github.com/Greysahy/CLDN.