Point Adversarial Self-Mining: A Simple Method for Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) accuracy is crucial for human-computer interaction.
• Existing FER methods often rely on complex architectures or specialized loss functions.
• There is a need for more effective and adaptive approaches to improve FER performance.

Purpose of the Study:

• To introduce a novel method, Point Adversarial Self Mining (PASM), for enhancing facial expression recognition accuracy.
• To simulate human learning processes by providing iterative guidance and adaptive learning materials.
• To improve network capability without designing task-specific architectures or loss functions.

Main Methods:

• PASM employs a point adversarial attack and a teacher network to identify informative positions for generating challenging learning samples.
• The method adaptively generates new learning materials considering sample statistics and teacher network capabilities.
• A teacher-student learning paradigm is utilized, where the student network becomes a teacher in subsequent iterations.

Main Results:

• Extensive experiments demonstrate the effectiveness of PASM in improving FER accuracy.
• The proposed method shows superior performance compared to existing state-of-the-art FER techniques.
• Iterative refinement through adaptive learning materials and teacher/student updates leads to enhanced network capability.

Conclusions:

• PASM offers a simple yet effective approach to boost facial expression recognition accuracy.
• The adaptive learning and iterative teacher-student update mechanism is key to the method's success.
• PASM represents a significant advancement in the field of facial expression recognition.