Robust and energy-efficient expression recognition based on improved deep ResNets.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• Deep Convolutional Neural Networks (CNNs), such as Residual Networks (ResNets), offer improved performance in AI tasks due to their depth.
• However, deeper networks like ResNets face challenges with high energy consumption, particularly on mobile devices.
• Robust and energy-efficient facial expression recognition is crucial for widespread AI application.

Purpose of the Study:

• To develop a facial expression recognition method that is both robust to noise and energy-efficient.
• To address the energy consumption limitations of deep ResNets in AI tasks.
• To explore the use of novel activation functions for improved performance on neuromorphic hardware.

Main Methods:

• Proposed an improved deep residual network (ResNet) architecture for facial expression recognition.
• Introduced and utilized the Noisy Softplus (NSP) activation function, replacing Rectified Linear Units (ReLU).
• Trained and evaluated an 18-layered ResNet with NSP on the Cohn-Kanade (CK+), Karolinska Directed Emotional Faces (KDEF), and GENKI-4K datasets.

Main Results:

• The NSP-based ResNet demonstrated superior anti-noise capabilities compared to ReLU-based ResNets.
• Facial expression recognition models trained with NSP exhibited significantly lower energy consumption.
• NSP-trained models were successfully implemented on ultra-low-power neuromorphic hardware.

Conclusions:

• The proposed NSP-enhanced ResNet offers a robust and energy-efficient solution for facial expression recognition.
• This approach facilitates the deployment of high-performance AI vision applications on power-constrained mobile and neuromorphic devices.
• The study highlights the potential of biologically plausible activation functions for efficient AI on specialized hardware.