H.5. Image Processing and Computer Vision
Khosro Rezaee
Abstract
Traditional Down syndrome identification often relies on professionals visually recognizing facial features, a method that can be subjective and inconsistent. This study introduces a hybrid deep learning (DL) model for automatically identifying Down syndrome in children's facial images, utilizing facial ...
Read More
Traditional Down syndrome identification often relies on professionals visually recognizing facial features, a method that can be subjective and inconsistent. This study introduces a hybrid deep learning (DL) model for automatically identifying Down syndrome in children's facial images, utilizing facial analysis techniques to enhance diagnostic accuracy and enable real-time detection. The model employs the MobileNetV2 architecture to address dataset bias and diversity issues while ensuring efficient feature extraction. The framework also integrates the structure with optimized Bidirectional Long Short-Term Memory (BiLSTM) to enhance feature classification. Trained and validated on facial images from children with Down syndrome and healthy controls from the Kaggle dataset, the model achieved 97.60% accuracy and 97.50% recall. The approach also integrates cloud and edge processing for efficient real-time analysis, offering adaptability to new images and conditions.
