Facial Expressions Recognition for Human-Robot Interaction Using Deep Convolutional Neural Networks with Rectified...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Human-robot interaction (HRI) systems require real-time processing for natural engagement.
• Deep convolutional neural networks (CNNs) offer powerful tools for visual perception tasks like face and emotion recognition.

Purpose of the Study:

• To develop an efficient end-to-end pipeline for real-time human-robot interaction using CNNs.
• To compare the performance of different CNN models for face recognition (FR) and facial expression recognition (FER).

Main Methods:

• An innovative end-to-end pipeline employing two optimized CNNs: one for FR and one for FER.
• Utilized Faster Region-based Convolutional Neural Network (Faster R-CNN) and Single Shot Detector (SSD) CNN for FR.
• Applied transfer learning and fine-tuning on VGG, Inception V3, and ResNet models for FER.

Main Results:

• Face detection models achieved high accuracy: Faster R-CNN at 97.8% and SSD Inception at 97.42%.
• ResNet yielded the highest FER training accuracy (90.14%), followed by VGG (87%) and Inception V3 (81%).
• Serialized CNNs improved performance by over 10%, with RAdam optimization enhancing generalization and accuracy by 3-4%.

Conclusions:

• The proposed pipeline enables real-time human-robot interaction through efficient face and facial expression recognition.
• Optimized CNN models provide a viable solution for enhancing the perception capabilities of social robots like NAO.