Facial Feature Segmentation for Masked Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face recognition systems struggle with identity recognition when facial features are partially masked, as the arrangement of output vectors from face recognition models is meaningless and can lead to errors due to the lack of correspondence with human-perceived facial features.
Innovation Solution
A training system and method that utilizes an encoder module, shared decoder module, and classification module to generate and update parameters based on feature tensors and classification labels, ensuring that the neural network focuses on facial features by incorporating a Gaussian mask and sigmoid function to enhance specific facial features like eyes, nose, and mouth, thereby generating vectors that retain meaningful facial information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a face recognition learning model generates output vectors through iterative learning, then the model can represent relevant facial information, but the arrangement of output vectors becomes meaningless and cannot correspond to human-perceived facial features
Solution Approach 1:
The patent segments the facial image into multiple feature regions (eyes, nose, mouth, etc.) and processes each region separately through the neural network. This segmentation allows the output vectors to correspond to specific facial features rather than being a meaningless aggregated representation, resolving the contradiction between information retention and meaningful arrangement.
Solution Approach 2:
The patent transforms the traditional single-vector output into a multi-dimensional structure where each dimension corresponds to a specific facial feature. By adding the dimension of feature localization, the system achieves meaningful vector arrangement that corresponds to human perception while preserving all facial information.
2Reliability
If facial features are partially masked by objects, then human recognition can still determine identity based on visible features like eyes, but current face recognition models generate errors because they cannot identify which vectors contain unmasked feature information
Solution Approach 1:
By segmenting the face into distinct feature regions, the system can independently evaluate the visibility and quality of each feature. When masking occurs, the system identifies which segments are obscured and relies more heavily on unmasked segments, enabling reliable recognition without requiring complex global mask detection mechanisms.
Solution Approach 2:
The patent applies different processing and weighting to different facial regions based on their local characteristics and visibility. Unmasked features receive higher weight in the recognition decision, while masked features are downweighted or excluded. This local quality approach enables the system to maintain high reliability under partial masking without adding significant complexity.
3Measurement precision
If the neural network focuses on extracting meaningful facial feature information, then recognition accuracy improves, but the model complexity and training requirements increase
Solution Approach 1:
The segmentation approach naturally guides the neural network to focus on specific facial features through localized processing. Each segment can be processed with simpler operations while the overall system achieves high feature extraction accuracy. This modular segmentation reduces the complexity burden compared to attempting to extract all features simultaneously in a monolithic architecture.
Data Source
AI summary
A training system, a training method, and a recognition system are provided. The training method is used to train a neural network module including: an encoder module, a shared decoder module, a synthesis module, and a classification module. The training method includes performing in a training epoch: repeatedly executing: taking a training image from a training set as an input image, obtaining a first loss based on training feature images of the training image and the feature images corresponding to the training image, and obtaining a second loss based on a classification marker of the training image and a classification generated by the classification module in correspondence with the training image; and updating first parameters and second parameters based on an average value of all the first losses and an average value of all the second losses obtained in the preceding step and an update algorithm.