Sparse Simultaneous Recurrent Deep Learning for Robust Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Facial expression recognition is complex, involving subtle muscle changes.
• Deep neural networks (DNNs) show promise but face challenges with depth and computational cost.
• Existing DNNs overlook the benefits of recurrent processing found in human vision.

Purpose of the Study:

• Propose a novel, biologically relevant sparse-deep simultaneous recurrent network (S-DSRN).
• Enhance facial expression recognition robustness and efficiency.
• Investigate the impact of dropout learning for feature sparsity.

Main Methods:

• Developed a sparse-deep simultaneous recurrent network (S-DSRN).
• Utilized dropout learning for inherent feature sparsity and regularization.
• Integrated S-DSRN with metric learning techniques.
• Implemented S-DSRN on a GPU for real-time performance.

Main Results:

• S-DSRN demonstrated superior performance accuracy over state-of-the-art feed-forward DNNs.
• The proposed network requires fewer parameters and exhibits reduced computational complexity.
• Dropout learning effectively induced sparsity and prevented overfitting.
• Combining S-DSRN with metric learning further boosted recognition performance.

Conclusions:

• The S-DSRN offers a robust and efficient solution for facial expression recognition.
• Biologically inspired recurrent architectures combined with sparse learning are highly effective.
• The S-DSRN is suitable for real-time applications, including those requiring GPU acceleration.