Facial Expression Embedding Model Decoupling Identity Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial expression embedding methods have relatively low accuracy due to the close coupling of identity information with expression features.
Innovation Solution
A training method for a facial expression embedding model that includes determining a sample set with three images and a label, and training a model with a full face embedding sub-model and an identity embedding sub-model, where the output is determined by the difference between the two sub-model outputs, decoupling identity and expression features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current embedding methods are used to represent facial expressions, then the implementation is simple, but the accuracy is relatively low due to close coupling of identity information with expression features
Solution Approach 1:
The patent divides the facial embedding task into two separate sub-models: an identity embedding sub-model that extracts identity features, and a full face embedding sub-model that extracts both identity and expression features. By segmenting the feature extraction process, the patent isolates expression features from identity information, thereby improving embedding accuracy while managing model complexity through modular architecture
Solution Approach 2:
The patent extracts identity features using a pretrained identity embedding sub-model and removes them from the full face embedding output. This extraction process isolates pure expression features by subtracting the identity component, directly addressing the accuracy problem caused by identity-expression coupling while maintaining a structured approach to model complexity
2Reliability
If identity information is included in facial expression embedding, then the model captures comprehensive facial information, but the expression feature extraction accuracy deteriorates due to coupling with identity information
Solution Approach 1:
The patent converts the harmful effect of identity information coupling into a benefit by using the identity embedding sub-model to explicitly capture and remove identity features. The identity information that previously caused interference is now harnessed to improve expression extraction reliability by serving as a subtractive component, transforming the problem into a solution
Solution Approach 2:
The patent introduces an intermediary subtraction operation that mediates between the full face embedding and identity embedding. This intermediary process selectively removes identity information while preserving expression features, thereby improving extraction reliability without complete loss of identity information, as the identity features are extracted and utilized in a controlled manner
Data Source
AI summary
The present disclosure provides a training method of a facial expression embedding model, a facial expression embedding method, and a facial expression embedding device. The method includes: determining a sample set, wherein each sample in the sample set includes three images and a sample label; and training the to-be-trained facial expression embedding model with the sample set, to obtain the trained facial expression embedding model, wherein the to-be-trained facial expression embedding model includes a to-be-trained full face embedding sub-model and a trained identity embedding sub-model, the trained facial expression embedding model includes a trained full face embedding sub-model and the trained identity embedding sub-model, and an output of the trained facial expression embedding model is determined by a difference between an output of the trained full face embedding sub-model and an output of the trained identity embedding sub-model.