Self-Difference Convolutional Neural Network for Facial Expression Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) faces challenges due to intra-class variations related to individual identities.
• Existing methods struggle to accurately classify expressions when subjects exhibit variations.

Purpose of the Study:

• To propose a Self-Difference Convolutional Network (SD-CNN) to mitigate intra-class variation in FER.
• To enhance the accuracy and efficiency of facial expression classification.

Main Methods:

• A conditional generative adversarial network synthesizes six typical facial expressions for a given subject.
• Six lightweight, difference-based Convolutional Neural Networks (DiffNets) compare deep features between test images and synthesized expressions.
• This self-difference approach alleviates intra-class variation by tightly clustering same-expression features.

Main Results:

• The SD-CNN achieved state-of-the-art accuracies of 99.7% on the CK+ dataset and 91.3% on the Oulu-CASIA dataset.
• The online processing model size is only 9.54 MB, enabling deployment on low-cost hardware.
• Significant alleviation of intra-class variation was observed.

Conclusions:

• The proposed SD-CNN effectively addresses the intra-class variation problem in facial expression recognition.
• The method demonstrates superior performance and efficiency, making it suitable for real-world applications.
• SD-CNN offers a promising solution for accurate and lightweight facial expression analysis.