Konsep Inti
A dual-branch adaptive distribution fusion framework is proposed to address the ambiguity problem in facial expression recognition by mining class distributions of emotions and adaptively fusing them with label distributions of samples.
Abstrak
The paper presents a novel multi-task framework, Ada-DF, for facial expression recognition (FER) that integrates label distribution generation as an auxiliary task. The framework consists of an auxiliary branch responsible for extracting label distributions of samples and a target branch for facial expression classification.
Key highlights:
- The auxiliary branch extracts label distributions of samples, which are then used to mine class distributions of emotions. These class distributions aim to exclude biases in the label distributions and capture the rich sentiment information behind each emotion.
- An adaptive distribution fusion module is proposed to balance the robustness of class distributions and the diversity of label distributions. Attention weights are used to adaptively fuse the two distributions, providing more accurate and comprehensive supervision for training the target branch.
- Extensive experiments on three real-world FER datasets (RAF-DB, AffectNet, and SFEW) demonstrate the effectiveness and robustness of the proposed Ada-DF framework, outperforming state-of-the-art methods.
- Detailed analysis reveals the significant contribution of label distribution extraction, class distribution mining, and adaptive distribution fusion in improving the FER performance.
- The framework has the potential for broader applicability in other deep learning-based tasks beyond FER.
Statistik
The RAF-DB dataset contains 29,672 real-world images, with a training set of 12,271 images and a test set of 2,478 images.
The AffectNet dataset contains over 1 million real-world images, with a training set of 287,651 images and a test set of 3,999 images.
The SFEW dataset contains 958 training images, 436 validation images, and 272 test images.
Kutipan
"Facial expression plays a pivotal role in human communication, which serves as a crucial medium for conveying emotions."
"Recent advancements in deep learning coupled with the availability of large-scale datasets have made great progress in FER, surpassing the performance of traditional methods."
"To address the ambiguity problem in FER, the label distribution learning (LDL) is introduced, which assigning different weights to all emotions."