Infant Facial Landmark Estimation Using Domain Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial landmark estimation models are inadequate for infant faces due to their rarity in datasets and challenging characteristics, such as small size and obscured appearances, leading to poor performance in identifying and tracking facial landmarks.
Innovation Solution
A new dataset, InfAnFace, with annotated infant faces and pose attributes is created, and state-of-the-art models using domain adaptation techniques are trained to improve facial landmark estimation, specifically using convolutional neural networks like HRNet and RetinaFace for better identification and tracking of infant facial features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If vision-based facial landmark estimation models are used, then adult face recognition is effective, but infant face estimation performance is poor due to small size and obscured appearance
Solution Approach 1:
The patent applies parameter changes by modifying the training parameters and architecture of convolutional neural networks to specialize in infant face characteristics. The model parameters are adjusted to handle the specific challenges of infant faces such as smaller feature sizes and different geometric proportions, enabling accurate landmark estimation that general adult face models cannot achieve.
Solution Approach 2:
The patent creates a dedicated infant face dataset and trains separate convolutional neural network models specifically for infant facial landmark estimation. This copying approach involves replicating the successful adult face estimation framework but with infant-specific training data and parameters, thereby adapting the model to the unique characteristics of infant faces without relying on generic adult models.
2Quantity of substance
If general facial landmark estimation models are trained on adult datasets, then training data availability is sufficient, but infant face performance is inadequate due to rarity in datasets
Solution Approach 1:
The patent performs preliminary action by creating a specialized infant face dataset before training the final model. This involves collecting, annotating, and preparing infant face images in advance, ensuring that the training data is ready and optimized for the specific task. This preliminary data preparation enables reliable model training despite the rarity of infant faces in general datasets.
Solution Approach 2:
The patent segments the facial landmark estimation task into separate models for infant and adult faces. Instead of training a single universal model, the system divides the training into distinct datasets and models, allowing each to be optimized for its specific domain. This segmentation enables reliable infant face estimation using infant-specific data while maintaining the ability to handle adult faces separately.
3Adaptability or versatility
If 3D scan-based infant face models are used, then pose variation capability is improved, but model accessibility is limited as only one such model exists and is not publicly available
Solution Approach 1:
The patent substitutes the mechanical 3D scanning system with a vision-based convolutional neural network approach. Instead of relying on physical 3D scans that require specialized equipment and are difficult to reproduce, the system uses standard 2D images processed through trained neural networks. This substitution maintains pose variation handling capability while dramatically improving accessibility and reproducibility through open-source software implementation.
Data Source
AI summary
Provided herein are methods and systems for identifying a face of an infant in an image including providing a computer comprising a processor and a memory trained with a set of training images and programmed with a convolutional neural network (CNN) model for identifying a face of an infant in a test image suspected of comprising an infant's face, wherein each image of the set of training images includes a plurality of facial landmark annotations and at least one pose attribute annotation, providing a test image suspected of comprising an image of an infant's face, and processing the test image using the computer, whereby the infant's face is identified in the test image.