Face Recognition Feature Decorrelation via Neural Network Factorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current age-invariant face recognition technologies face challenges due to the correlation between identity and age features, leading to unreliable recognition results, as age information is often carried by identity features.
Innovation Solution
A model training method that decorrelates identity and age features using a feature extraction module implemented with a neural network, where a correlation coefficient is determined and used to perform decorrelated training, reducing the correlation between these features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-task learning of identity discrimination and age discrimination is used, then face recognition capability is improved, but recognition reliability deteriorates due to correlation between identity and age features
Solution Approach 1:
The patent segments the identity feature into two independent components: an age-invariant identity feature (unique to the individual) and an age feature (related to chronological age). This segmentation is achieved through a factorization module that decomposes the original identity feature representation. By separating these features, the system can use the age-invariant component for recognition while ignoring the age component, thus resolving the contradiction between recognition accuracy and reliability.
Solution Approach 2:
The patent extracts and removes the age feature component from the identity feature representation. The factorization module explicitly identifies and separates the age-related information from the identity-related information. During recognition, only the age-invariant identity feature is used, effectively taking out the harmful age correlation that would otherwise degrade recognition reliability.
2Adaptability or versatility
If identity feature carries age information, then age discrimination capability is improved, but face recognition reliability deteriorates
Solution Approach 1:
The patent segments the identity feature into two independent components: an age-invariant identity feature (unique to the individual) and an age feature (related to chronological age). This segmentation is achieved through a factorization module that decomposes the original identity feature representation. By separating these features, the system can use the age-invariant component for recognition while ignoring the age component, thus resolving the contradiction between recognition accuracy and reliability.
Solution Approach 2:
The patent extracts and removes the age feature component from the identity feature representation. The factorization module explicitly identifies and separates the age-related information from the identity-related information. During recognition, only the age-invariant identity feature is used, effectively taking out the harmful age correlation that would otherwise degrade recognition reliability.
Data Source
AI summary
A face recognition method includes: extracting a first identity feature of a first face image by using a feature extraction module, and extracting a second identity feature of a second face image by using the feature extraction module, wherein the feature extraction module is implemented by using a neural network, and pre-trained in a manner such that a correlation coefficient of training batch data is obtained based on an identity feature and an age feature of a sample face image in the training batch data, and decorrelated training of the identity feature and the age feature is performed on the feature extraction module based on the correlation coefficient; and performing a face recognition based on determining a similarity between faces in the first face image and the second face image according to the first identity feature and the second identity feature.