Neural Network Architecture for Facial Movement Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning architectures for analyzing facial movements to diagnose medical conditions, such as tardive dyskinesia and Parkinson's disease, often produce inaccurate and unreliable results due to their complexity and limited training data.
Innovation Solution
A computer-implemented method that segments a video into sequences corresponding to different facial states, tracks the displacement of groups of facial landmarks, and uses multiple specialized neural networks to determine landmark features and sequence scores, which are then combined to obtain a patient score indicative of a disease state.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single complex machine learning architecture is used to analyze facial movements, then the system can process all facial states in one model, but the accuracy and reliability of disease state determination deteriorates due to complexity and limited training data
Solution Approach 1:
The patent divides the facial analysis task into multiple specialized neural networks, each dedicated to analyzing specific facial states (e.g., mouth movements, eye movements, facial expressions). This segmentation allows each network to specialize in particular movement patterns, improving reliability for disease state determination while maintaining the ability to process all facial states through the coordinated work of multiple networks
2Measurement precision
If multiple specialized neural networks are used to analyze different facial landmark groups, then the measurement precision of displacement data improves, but the device complexity increases
Solution Approach 1:
The patent applies local quality by creating different neural networks optimized for specific facial landmark groups (e.g., eyes, mouth, nose). Each network is trained to precisely measure displacement patterns for its designated facial region, improving measurement precision for each local area while the overall system complexity is managed through modular organization of these specialized components
Data Source
AI summary
A video is segmented into a plurality of sequences corresponding to different facial states performed by a patient in the video. For each sequence, displacement of a plurality of groups of landmarks of a face of the patient is tracked, to obtain, for each group of the plurality of groups, one or more displacement measures characterizing positions of the landmarks of the group. The one or more displacement measures corresponding to each group are provided into a corresponding neural network, to obtain a landmark feature. The neural networks corresponding to each group are different from one another. A sequence score for the sequence is determined based on a plurality of landmark features corresponding to the groups. A plurality of sequence scores are provided into a machine learning component, to obtain a patient score. A disease state of the patient is determined based on the patient score.