Training and Profiling a Pediatric Facial Expression Classifier for Children on Mobile Devices: Machine Learning...

Area of Science:

• Computer Vision
• Machine Learning
• Developmental Psychology

Background:

• Facial expression recognition (FER) on mobile devices can aid individuals with conditions like autism.
• Existing FER models are computationally intensive, limiting smartphone application.
• Advancements are needed for efficient and accessible FER tools.

Purpose of the Study:

• To optimize state-of-the-art facial expression classifiers for mobile device efficiency and performance.
• To evaluate the impact of training data (children vs. adults) on classifier accuracy.
• To assess model performance across diverse ethnic groups.

Main Methods:

• Utilized 12 public datasets and crowdsourced videos for classifier training.
• Tested 5 convolutional neural network architectures (e.g., MobileNetV3, EfficientNetB0).
• Applied optimization techniques like weight pruning and quantize-aware training; evaluated on a Motorola Moto G6.

Main Results:

• A MobileNetV3-Large model achieved 65.78% accuracy and 90ms inference on a Moto G6.
• This optimized model approached state-of-the-art performance with significantly fewer parameters.
• Models trained on children outperformed those trained on adults; performance varied across ethnicities.

Conclusions:

• Specialized design and optimization enable lightweight, high-performance mobile FER.
• A 'data shift' exists between child and adult facial expressions, favoring child-trained models.
• Underrepresented ethnic groups showed performance disparities, necessitating further research for equitable AI in healthcare.