Multitask, Multilabel, and Multidomain Learning With Convolutional Networks for Emotion Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automated emotion recognition from facial images is challenging due to pose, orientation, and viewpoint variations.
• Current deep learning models struggle with these real-world variations and the cost of labeled data.
• Existing methods for emotion recognition are limited in handling diverse and uncontrolled facial data.

Purpose of the Study:

• To propose a novel multitask learning loss function for improved emotion recognition.
• To enhance feature representation by sharing commonalities with related tasks like facial action unit detection.
• To address the challenge of learning from heterogeneously labeled datasets in uncontrolled environments.

Main Methods:

• Applied a multitask learning loss function to jointly train emotion recognition and facial action unit detection models.
• Developed a shared feature representation to leverage related task information.
• Validated the approach on three diverse datasets acquired in non-controlled environments.

Main Results:

• Demonstrated that jointly learning emotion recognition with facial action unit detection significantly improves performance.
• The proposed loss function effectively handles heterogeneously labeled data across tasks.
• The model showed robust performance in recognizing emotions in challenging, real-world conditions.

Conclusions:

• Multitask learning, particularly with facial action units, offers a promising direction for robust emotion recognition.
• The developed loss function advances multitask learning for heterogeneously labeled data.
• This approach provides a more effective solution for automated emotion recognition in the wild.