Reliable Crowdsourcing and Deep Locality-Preserving Learning for Unconstrained Facial Expression Recognition

Area of Science:

• Computer Vision
• Affective Computing
• Machine Learning

Background:

• Facial expression analysis is crucial for human interaction.
• Existing facial expression databases often lack real-world variability.
• Posed expressions in controlled settings limit understanding of natural human emotions.

Purpose of the Study:

• To introduce a large-scale, diverse facial expression database captured in real-world conditions.
• To develop and validate a deep learning model for robust facial expression recognition.
• To investigate the complexities of real-world emotions, including compound and mixed expressions.

Main Methods:

• Creation of the Real-world Affective Face Database (RAF-DB) with ~30,000 diverse, uncontrolled facial images.
• Development of an expectation-maximization algorithm for reliable emotion label estimation.
• Proposal of a deep locality-preserving convolutional neural network (DLP-CNN) for expression recognition.

Main Results:

• RAF-DB contains diverse facial expressions with compound and mixed emotions.
• Real-world facial action units are more varied than laboratory-controlled ones.
• The proposed DLP-CNN significantly outperforms state-of-the-art methods on multiple benchmark datasets.

Conclusions:

• The RAF-DB provides a valuable resource for studying natural facial expressions.
• DLP-CNN offers enhanced discriminative power for in-the-wild expression recognition.
• The findings highlight the need for more realistic datasets and advanced models in affective computing.