Progressive ShallowNet for large scale dynamic and spontaneous facial behaviour analysis in children

Area of Science:

• Computer Science
• Psychology
• Neuroscience

Background:

• COVID-19 pandemic exacerbated social isolation, impacting mental health and social competence in both adults and children.
• Deficits in recognizing facial emotions in children can lead to impaired social functioning and long-term implications.
• Existing research predominantly focuses on adult emotion recognition, neglecting the critical developmental period in childhood.

Purpose of the Study:

• To develop an efficient and lightweight deep learning model for spontaneous facial behavior recognition in children.
• To address the limitations of existing models in handling smaller datasets and preventing overfitting.
• To improve early identification of emotional deficits in children for better social functioning.

Main Methods:

• A novel progressive lightweight shallow learning architecture, Progressive ShallowNet, inspired by pyramidal cells in the cerebral cortex.
• Efficient utilization of skip-connections to manage gradient flow and gradually increase residual path depth.
• Extensive experimentation on benchmark datasets for facial behavior analysis in children.

Main Results:

• Progressive ShallowNet demonstrated significant performance gains in spontaneous facial behavior recognition compared to existing methods.
• The model effectively explores a larger feature space while mitigating overfitting issues common with smaller datasets.
• The architecture's design, limiting the residual path locally, enhances robustness against perturbations.

Conclusions:

• The proposed Progressive ShallowNet offers an effective solution for emotion recognition in children, crucial for early intervention.
• This advancement can aid in preventing social functioning impairments stemming from deficits in emotional competence.
• The study highlights the importance of specialized models for child-specific emotion recognition tasks.