LEMON: A Lightweight Facial Emotion Recognition System for Assistive Robotics Based on Dilated Residual Convolutional...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Social intelligence systems in assistive robotics require real-time emotion recognition (ER) for empathic user interaction.
• Low-cost robots often have limited memory and processing power, posing challenges for complex ER models.
• A trade-off between accuracy and model complexity is necessary for effective ER in resource-constrained assistive robots.

Purpose of the Study:

• To propose a compact and robust emotion recognition service for assistive robotics, named Lightweight EMotion recognitiON (LEMON).
• To develop a deep learning model that balances accuracy and computational efficiency for facial expression recognition on low-cost hardware.

Main Methods:

• Utilized image processing, computer vision, and deep learning algorithms for facial expression recognition.
• Developed a deep learning model based on Residual Convolutional Neural Networks combined with Dilated and Standard Convolution Layers.
• Employed Exponential Linear Unit (ELU) activation function to enhance model stability and mitigate the dying ReLU problem.

Main Results:

• The proposed LEMON model features a significantly reduced parameter count (1.6 Million) compared to other approaches.
• Dilated convolutions were used to expand receptive fields efficiently, preserving resolution and reducing computational/memory costs.
• The model achieved comparable results to existing methods across multiple datasets, demonstrating its effectiveness and efficiency.

Conclusions:

• The LEMON service offers a viable solution for integrating effective emotion recognition into cost-effective assistive robots.
• The compact design and robust performance of LEMON address the hardware limitations of low-cost robotic platforms.
• This research contributes to the advancement of empathic human-robot interaction through efficient AI.