Full-convolution Siamese network algorithm under deep learning used in tracking of facial video image in newborns

Area of Science:

• Computer Vision
• Neonatal Care
• Machine Learning

Background:

• Current neonatal facial recognition technology lacks accuracy.
• Accurate tracking of infant facial features is crucial for pain assessment and emotional evaluation.
• Deep learning approaches offer potential for improving neonatal facial image analysis.

Purpose of the Study:

• To optimize the full-convolution Siamese network (SiamFC) for neonatal facial video image tracking.
• To develop an improved algorithm for accurate recognition of neonatal pain and emotions.
• To enhance the reliability of automated infant facial analysis.

Main Methods:

• Optimization of the SiamFC algorithm for deep learning-based facial recognition.
• Construction of a newborn facial video image tracking (FVIT) model using SiamFC and an attention mechanism.
• Development of a newborn face database for model evaluation.

Main Results:

• The improved algorithm achieved an accuracy of 0.889, outperforming other models by 0.036.
• The area under the curve (AUC) for success rate reached 0.748, an improvement of 0.075.
• The FVIT model demonstrated robust performance in tracking occluded faces, expressions, and scale variations.

Conclusions:

• The optimized SiamFC-based FVIT model significantly enhances accuracy and success rate in neonatal facial image tracking.
• This approach provides a reliable method for capturing and analyzing infant facial features, aiding pain assessment.
• The study establishes an experimental foundation for advanced facial recognition and pain evaluation in newborns.