Improving facial expression recognition for autism with IDenseNet-RCAformer under occlusions

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Facial expression recognition (FER) is vital for social interaction, but accuracy decreases with occlusions like glasses or facial hair.
• Individuals with autism spectrum disorder (ASD) often face challenges in understanding and expressing emotions, making accurate FER crucial for them.
• Existing FER systems struggle with partial facial occlusions, particularly in the context of ASD.

Purpose of the Study:

• To develop an advanced framework for facial expression recognition (FER) that effectively handles partial occlusions in individuals with autism spectrum disorder (ASD).
• To improve the accuracy and efficiency of FER systems for a neurodevelopmental disorder characterized by social interaction impairments.
• To introduce a novel deep learning model, the Improved DenseNet-based Residual Cross-Attention Transformer (IDenseNet-RCAformer), for robust FER in challenging conditions.

Main Methods:

• A novel IDenseNet-RCAformer system was proposed, integrating Inception-ResNet-V2 for local features and Cross-Attention Transformer for global features.
• Features were extracted and fused using the FusionNet method to enhance training speed and precision.
• Preprocessing techniques and an argmax function for heatmap-based landmark prediction were employed to optimize recognition efficiency.

Main Results:

• The proposed IDenseNet-RCAformer system demonstrated superior performance in facial expression recognition for partially occluded faces in autism patients.
• The system achieved a high accuracy rate of 98.95% on four facial expression datasets.
• Experimental results indicated that the IDenseNet-RCAformer significantly outperformed previous facial expression recognition frameworks.

Conclusions:

• The IDenseNet-RCAformer system offers a significant advancement in addressing the challenge of facial expression recognition with occlusions in individuals with autism spectrum disorder.
• The proposed method provides a more accurate and efficient solution compared to existing FER frameworks for this specific population.
• This research contributes to better understanding and potentially aiding social interaction for individuals with ASD through improved emotion recognition technology.